+
+ This program is free software: you can redistribute it and/or modify
+ it under the terms of the GNU Affero General Public License as published
+ by the Free Software Foundation, either version 3 of the License, or
+ (at your option) any later version.
+
+ This program is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU Affero General Public License for more details.
+
+ You should have received a copy of the GNU Affero General Public License
+ along with this program. If not, see .
+
+Also add information on how to contact you by electronic and paper mail.
+
+ If your software can interact with users remotely through a computer
+network, you should also make sure that it provides a way for users to
+get its source. For example, if your program is a web application, its
+interface could display a "Source" link that leads users to an archive
+of the code. There are many ways you could offer source, and different
+solutions will be better for different programs; see section 13 for the
+specific requirements.
+
+ You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU AGPL, see
+.
diff --git a/README.md b/README.md
index ffdbe63f08fbe333ee373f78666574eeed514517..7a5c8aec33f74b88f8356ff0c8eab8996224d5d3 100644
--- a/README.md
+++ b/README.md
@@ -1,13 +1,199 @@
----
-title: RVC UI
-emoji: 🏢
-colorFrom: red
-colorTo: gray
-sdk: gradio
-sdk_version: 4.39.0
-app_file: app.py
-pinned: false
-license: mit
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+An easy-to-use voice conversion framework based on VITS.
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+[**FAQ (Frequently Asked Questions)**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/FAQ-(Frequently-Asked-Questions))
+
+[**English**](./README.md) | [**中文简体**](./docs/cn/README.cn.md) | [**日本語**](./docs/jp/README.ja.md) | [**한국어**](./docs/kr/README.ko.md) ([**韓國語**](./docs/kr/README.ko.han.md)) | [**Français**](./docs/fr/README.fr.md) | [**Türkçe**](./docs/tr/README.tr.md) | [**Português**](./docs/pt/README.pt.md)
+
+
+
+> The base model is trained using nearly 50 hours of high-quality open-source VCTK training set. Therefore, there are no copyright concerns, please feel free to use.
+
+> Please look forward to the base model of RVCv3 with larger parameters, larger dataset, better effects, basically flat inference speed, and less training data required.
+
+> There's a [one-click downloader](https://github.com/fumiama/RVC-Models-Downloader) for models/integration packages/tools. Welcome to try.
+
+| Training and inference Webui |
+| :--------: |
+|  |
+
+| Real-time voice changing GUI |
+| :---------: |
+|  |
+
+## Features:
++ Reduce tone leakage by replacing the source feature to training-set feature using top1 retrieval;
++ Easy + fast training, even on poor graphics cards;
++ Training with a small amounts of data (>=10min low noise speech recommended);
++ Model fusion to change timbres (using ckpt processing tab->ckpt merge);
++ Easy-to-use WebUI;
++ UVR5 model to quickly separate vocals and instruments;
++ High-pitch Voice Extraction Algorithm [InterSpeech2023-RMVPE](#Credits) to prevent a muted sound problem. Provides the best results (significantly) and is faster with lower resource consumption than Crepe_full;
++ AMD/Intel graphics cards acceleration supported;
++ Intel ARC graphics cards acceleration with IPEX supported.
+
+Check out our [Demo Video](https://www.bilibili.com/video/BV1pm4y1z7Gm/) here!
+
+## Environment Configuration
+### Python Version Limitation
+> It is recommended to use conda to manage the Python environment.
+
+> For the reason of the version limitation, please refer to this [bug](https://github.com/facebookresearch/fairseq/issues/5012).
+
+```bash
+python --version # 3.8 <= Python < 3.11
+```
+
+### Linux/MacOS One-click Dependency Installation & Startup Script
+By executing `run.sh` in the project root directory, you can configure the `venv` virtual environment, automatically install the required dependencies, and start the main program with one click.
+```bash
+sh ./run.sh
+```
+
+### Manual Installation of Dependencies
+1. Install `pytorch` and its core dependencies, skip if already installed. Refer to: https://pytorch.org/get-started/locally/
+ ```bash
+ pip install torch torchvision torchaudio
+ ```
+2. If you are using Nvidia Ampere architecture (RTX30xx) in Windows, according to the experience of #21, you need to specify the cuda version corresponding to pytorch.
+ ```bash
+ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+ ```
+3. Install the corresponding dependencies according to your own graphics card.
+- Nvidia GPU
+ ```bash
+ pip install -r requirements/main.txt
+ ```
+- AMD/Intel GPU
+ ```bash
+ pip install -r requirements/dml.txt
+ ```
+- AMD ROCM (Linux)
+ ```bash
+ pip install -r requirements/amd.txt
+ ```
+- Intel IPEX (Linux)
+ ```bash
+ pip install -r requirements/ipex.txt
+ ```
+
+## Preparation of Other Files
+### 1. Assets
+> RVC requires some models located in the `assets` folder for inference and training.
+#### Check/Download Automatically (Default)
+> By default, RVC can automatically check the integrity of the required resources when the main program starts.
+
+> Even if the resources are not complete, the program will continue to start.
+
+- If you want to download all resources, please add the `--update` parameter.
+- If you want to skip the resource integrity check at startup, please add the `--nocheck` parameter.
+
+#### Download Manually
+> All resource files are located in [Hugging Face space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+> You can find some scripts to download them in the `tools` folder
+
+> You can also use the [one-click downloader](https://github.com/fumiama/RVC-Models-Downloader) for models/integration packages/tools
+
+Below is a list that includes the names of all pre-models and other files required by RVC.
+
+- ./assets/hubert/hubert_base.pt
+ ```bash
+ rvcmd assets/hubert # RVC-Models-Downloader command
+ ```
+- ./assets/pretrained
+ ```bash
+ rvcmd assets/v1 # RVC-Models-Downloader command
+ ```
+- ./assets/uvr5_weights
+ ```bash
+ rvcmd assets/uvr5 # RVC-Models-Downloader command
+ ```
+If you want to use the v2 version of the model, you need to download additional resources in
+
+- ./assets/pretrained_v2
+ ```bash
+ rvcmd assets/v2 # RVC-Models-Downloader command
+ ```
+
+### 2. Download the required files for the rmvpe vocal pitch extraction algorithm
+
+If you want to use the latest RMVPE vocal pitch extraction algorithm, you need to download the pitch extraction model parameters and place them in `assets/rmvpe`.
+
+- [rmvpe.pt](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+#### Download DML environment of RMVPE (optional, for AMD/Intel GPU)
+
+- [rmvpe.onnx](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+### 3. AMD ROCM (optional, Linux only)
+
+If you want to run RVC on a Linux system based on AMD's ROCM technology, please first install the required drivers [here](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html).
+
+If you are using Arch Linux, you can use pacman to install the required drivers.
+````
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+````
+For some models of graphics cards, you may need to configure the following environment variables (such as: RX6700XT).
+````
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+````
+Also, make sure your current user is in the `render` and `video` user groups.
+````
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+````
+## Getting Started
+### Direct Launch
+Use the following command to start the WebUI.
+```bash
+python web.py
+```
+### Linux/MacOS
+```bash
+./run.sh
+```
+### For I-card users who need to use IPEX technology (Linux only)
+```bash
+source /opt/intel/oneapi/setvars.sh
+./run.sh
+```
+### Using the Integration Package (Windows Users)
+Download and unzip `RVC-beta.7z`. After unzipping, double-click `go-web.bat` to start the program with one click.
+```bash
+rvcmd packs/general/latest # RVC-Models-Downloader command
+```
+
+## Credits
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
++ [Vocal pitch extraction:RMVPE](https://github.com/Dream-High/RMVPE)
+ + The pretrained model is trained and tested by [yxlllc](https://github.com/yxlllc/RMVPE) and [RVC-Boss](https://github.com/RVC-Boss).
+
+## Thanks to all contributors for their efforts
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
diff --git a/assets/hubert/.gitignore b/assets/hubert/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/hubert/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/assets/hubert/hubert_base.pt b/assets/hubert/hubert_base.pt
new file mode 100644
index 0000000000000000000000000000000000000000..72f47ab58564f01d5cc8b05c63bdf96d944551ff
--- /dev/null
+++ b/assets/hubert/hubert_base.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f54b40fd2802423a5643779c4861af1e9ee9c1564dc9d32f54f20b5ffba7db96
+size 189507909
diff --git a/assets/indices/.gitignore b/assets/indices/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/indices/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/assets/pretrained/.gitignore b/assets/pretrained/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/pretrained/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/assets/pretrained_v2/.gitignore b/assets/pretrained_v2/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/pretrained_v2/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/assets/rmvpe/.gitignore b/assets/rmvpe/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/rmvpe/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/assets/rmvpe/rmvpe.pt b/assets/rmvpe/rmvpe.pt
new file mode 100644
index 0000000000000000000000000000000000000000..6362f060846875c3b5d7012adea5f97e47305e7e
--- /dev/null
+++ b/assets/rmvpe/rmvpe.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d62215f4306e3ca278246188607209f09af3dc77ed4232efdd069798c4ec193
+size 181184272
diff --git a/assets/uvr5_weights/.gitignore b/assets/uvr5_weights/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/assets/uvr5_weights/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/configs/__init__.py b/configs/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e9ab5e6d85936c4ff37838b5d66bcd4c4ce2855b
--- /dev/null
+++ b/configs/__init__.py
@@ -0,0 +1 @@
+from .config import singleton_variable, Config, CPUConfig
diff --git a/configs/config.json b/configs/config.json
new file mode 100644
index 0000000000000000000000000000000000000000..1aaa63fd3b7253a1b4271b0ca4e2dedafded559a
--- /dev/null
+++ b/configs/config.json
@@ -0,0 +1,21 @@
+{
+ "pth_path": "",
+ "index_path": "",
+ "sg_hostapi": "MME",
+ "sg_wasapi_exclusive": false,
+ "sg_input_device": "",
+ "sg_output_device": "",
+ "sr_type": "sr_device",
+ "threhold": -60.0,
+ "pitch": 12.0,
+ "formant": 0.0,
+ "rms_mix_rate": 0.5,
+ "index_rate": 0.0,
+ "block_time": 0.15,
+ "crossfade_length": 0.08,
+ "extra_time": 2.0,
+ "n_cpu": 4.0,
+ "use_jit": false,
+ "use_pv": false,
+ "f0method": "fcpe"
+}
\ No newline at end of file
diff --git a/configs/config.py b/configs/config.py
new file mode 100644
index 0000000000000000000000000000000000000000..0e8057c8b2495865c9d607c1e1114da7e57f9480
--- /dev/null
+++ b/configs/config.py
@@ -0,0 +1,259 @@
+import argparse
+import os
+import sys
+import json
+import shutil
+from multiprocessing import cpu_count
+
+import torch
+
+# TODO: move device selection into rvc
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+version_config_list = [
+ "v1/32k.json",
+ "v1/40k.json",
+ "v1/48k.json",
+ "v2/48k.json",
+ "v2/32k.json",
+]
+
+
+def singleton_variable(func):
+ def wrapper(*args, **kwargs):
+ if wrapper.instance is None:
+ wrapper.instance = func(*args, **kwargs)
+ return wrapper.instance
+
+ wrapper.instance = None
+ return wrapper
+
+
+@singleton_variable
+class Config:
+ def __init__(self):
+ self.device = "cuda:0"
+ self.is_half = True
+ self.use_jit = False
+ self.n_cpu = 0
+ self.gpu_name = None
+ self.json_config = self.load_config_json()
+ self.gpu_mem = None
+ (
+ self.python_cmd,
+ self.listen_port,
+ self.global_link,
+ self.noparallel,
+ self.noautoopen,
+ self.dml,
+ self.nocheck,
+ self.update,
+ ) = self.arg_parse()
+ self.instead = ""
+ self.preprocess_per = 3.7
+ self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
+
+ @staticmethod
+ def load_config_json() -> dict:
+ d = {}
+ for config_file in version_config_list:
+ p = f"configs/inuse/{config_file}"
+ if not os.path.exists(p):
+ shutil.copy(f"configs/{config_file}", p)
+ with open(f"configs/inuse/{config_file}", "r") as f:
+ d[config_file] = json.load(f)
+ return d
+
+ @staticmethod
+ def arg_parse() -> tuple:
+ exe = sys.executable or "python"
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--port", type=int, default=7865, help="Listen port")
+ parser.add_argument("--pycmd", type=str, default=exe, help="Python command")
+ parser.add_argument(
+ "--global_link", action="store_true", help="Generate a global proxy link"
+ )
+ parser.add_argument(
+ "--noparallel", action="store_true", help="Disable parallel processing"
+ )
+ parser.add_argument(
+ "--noautoopen",
+ action="store_true",
+ help="Do not open in browser automatically",
+ )
+ parser.add_argument(
+ "--dml",
+ action="store_true",
+ help="torch_dml",
+ )
+ parser.add_argument(
+ "--nocheck", action="store_true", help="Run without checking assets"
+ )
+ parser.add_argument(
+ "--update", action="store_true", help="Update to latest assets"
+ )
+ cmd_opts = parser.parse_args()
+
+ cmd_opts.port = cmd_opts.port if 0 <= cmd_opts.port <= 65535 else 7865
+
+ return (
+ cmd_opts.pycmd,
+ cmd_opts.port,
+ cmd_opts.global_link,
+ cmd_opts.noparallel,
+ cmd_opts.noautoopen,
+ cmd_opts.dml,
+ cmd_opts.nocheck,
+ cmd_opts.update,
+ )
+
+ # has_mps is only available in nightly pytorch (for now) and MasOS 12.3+.
+ # check `getattr` and try it for compatibility
+ @staticmethod
+ def has_mps() -> bool:
+ if not torch.backends.mps.is_available():
+ return False
+ try:
+ torch.zeros(1).to(torch.device("mps"))
+ return True
+ except Exception:
+ return False
+
+ @staticmethod
+ def has_xpu() -> bool:
+ if hasattr(torch, "xpu") and torch.xpu.is_available():
+ return True
+ else:
+ return False
+
+ def use_fp32_config(self):
+ for config_file in version_config_list:
+ self.json_config[config_file]["train"]["fp16_run"] = False
+ with open(f"configs/inuse/{config_file}", "r") as f:
+ strr = f.read().replace("true", "false")
+ with open(f"configs/inuse/{config_file}", "w") as f:
+ f.write(strr)
+ logger.info("overwrite " + config_file)
+ self.preprocess_per = 3.0
+ logger.info("overwrite preprocess_per to %d" % (self.preprocess_per))
+
+ def device_config(self):
+ if torch.cuda.is_available():
+ if self.has_xpu():
+ self.device = self.instead = "xpu:0"
+ self.is_half = True
+ i_device = int(self.device.split(":")[-1])
+ self.gpu_name = torch.cuda.get_device_name(i_device)
+ if (
+ ("16" in self.gpu_name and "V100" not in self.gpu_name.upper())
+ or "P40" in self.gpu_name.upper()
+ or "P10" in self.gpu_name.upper()
+ or "1060" in self.gpu_name
+ or "1070" in self.gpu_name
+ or "1080" in self.gpu_name
+ ):
+ logger.info("Found GPU %s, force to fp32", self.gpu_name)
+ self.is_half = False
+ self.use_fp32_config()
+ else:
+ logger.info("Found GPU %s", self.gpu_name)
+ self.gpu_mem = int(
+ torch.cuda.get_device_properties(i_device).total_memory
+ / 1024
+ / 1024
+ / 1024
+ + 0.4
+ )
+ if self.gpu_mem <= 4:
+ self.preprocess_per = 3.0
+ elif self.has_mps():
+ logger.info("No supported Nvidia GPU found")
+ self.device = self.instead = "mps"
+ self.is_half = False
+ self.use_fp32_config()
+ else:
+ logger.info("No supported Nvidia GPU found")
+ self.device = self.instead = "cpu"
+ self.is_half = False
+ self.use_fp32_config()
+
+ if self.n_cpu == 0:
+ self.n_cpu = cpu_count()
+
+ if self.is_half:
+ # 6G显存配置
+ x_pad = 3
+ x_query = 10
+ x_center = 60
+ x_max = 65
+ else:
+ # 5G显存配置
+ x_pad = 1
+ x_query = 6
+ x_center = 38
+ x_max = 41
+
+ if self.gpu_mem is not None and self.gpu_mem <= 4:
+ x_pad = 1
+ x_query = 5
+ x_center = 30
+ x_max = 32
+ if self.dml:
+ logger.info("Use DirectML instead")
+ import torch_directml
+
+ self.device = torch_directml.device(torch_directml.default_device())
+ self.is_half = False
+ else:
+ if self.instead:
+ logger.info(f"Use {self.instead} instead")
+ logger.info(
+ "Half-precision floating-point: %s, device: %s"
+ % (self.is_half, self.device)
+ )
+ return x_pad, x_query, x_center, x_max
+
+
+@singleton_variable
+class CPUConfig:
+ def __init__(self):
+ self.device = "cpu"
+ self.is_half = False
+ self.use_jit = False
+ self.n_cpu = 1
+ self.gpu_name = None
+ self.json_config = self.load_config_json()
+ self.gpu_mem = None
+ self.instead = "cpu"
+ self.preprocess_per = 3.7
+ self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
+
+ @staticmethod
+ def load_config_json() -> dict:
+ d = {}
+ for config_file in version_config_list:
+ with open(f"configs/{config_file}", "r") as f:
+ d[config_file] = json.load(f)
+ return d
+
+ def use_fp32_config(self):
+ for config_file in version_config_list:
+ self.json_config[config_file]["train"]["fp16_run"] = False
+ self.preprocess_per = 3.0
+
+ def device_config(self):
+ self.use_fp32_config()
+
+ if self.n_cpu == 0:
+ self.n_cpu = cpu_count()
+
+ # 5G显存配置
+ x_pad = 1
+ x_query = 6
+ x_center = 38
+ x_max = 41
+
+ return x_pad, x_query, x_center, x_max
diff --git a/configs/inuse/.gitignore b/configs/inuse/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..419423674c957a466a985a929c6b846494643f57
--- /dev/null
+++ b/configs/inuse/.gitignore
@@ -0,0 +1,4 @@
+*
+!.gitignore
+!v1
+!v2
diff --git a/configs/inuse/v1/.gitignore b/configs/inuse/v1/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/configs/inuse/v1/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/configs/inuse/v2/.gitignore b/configs/inuse/v2/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/configs/inuse/v2/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/configs/v1/32k.json b/configs/v1/32k.json
new file mode 100644
index 0000000000000000000000000000000000000000..d5f16d691ed798f4c974b431167c36269b2ce7d2
--- /dev/null
+++ b/configs/v1/32k.json
@@ -0,0 +1,46 @@
+{
+ "train": {
+ "log_interval": 200,
+ "seed": 1234,
+ "epochs": 20000,
+ "learning_rate": 1e-4,
+ "betas": [0.8, 0.99],
+ "eps": 1e-9,
+ "batch_size": 4,
+ "fp16_run": true,
+ "lr_decay": 0.999875,
+ "segment_size": 12800,
+ "init_lr_ratio": 1,
+ "warmup_epochs": 0,
+ "c_mel": 45,
+ "c_kl": 1.0
+ },
+ "data": {
+ "max_wav_value": 32768.0,
+ "sampling_rate": 32000,
+ "filter_length": 1024,
+ "hop_length": 320,
+ "win_length": 1024,
+ "n_mel_channels": 80,
+ "mel_fmin": 0.0,
+ "mel_fmax": null
+ },
+ "model": {
+ "inter_channels": 192,
+ "hidden_channels": 192,
+ "filter_channels": 768,
+ "n_heads": 2,
+ "n_layers": 6,
+ "kernel_size": 3,
+ "p_dropout": 0,
+ "resblock": "1",
+ "resblock_kernel_sizes": [3,7,11],
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+ "upsample_rates": [10,4,2,2,2],
+ "upsample_initial_channel": 512,
+ "upsample_kernel_sizes": [16,16,4,4,4],
+ "use_spectral_norm": false,
+ "gin_channels": 256,
+ "spk_embed_dim": 109
+ }
+}
diff --git a/configs/v1/40k.json b/configs/v1/40k.json
new file mode 100644
index 0000000000000000000000000000000000000000..4ffc87b9e9725fcd59d81a68d41a61962213b777
--- /dev/null
+++ b/configs/v1/40k.json
@@ -0,0 +1,46 @@
+{
+ "train": {
+ "log_interval": 200,
+ "seed": 1234,
+ "epochs": 20000,
+ "learning_rate": 1e-4,
+ "betas": [0.8, 0.99],
+ "eps": 1e-9,
+ "batch_size": 4,
+ "fp16_run": true,
+ "lr_decay": 0.999875,
+ "segment_size": 12800,
+ "init_lr_ratio": 1,
+ "warmup_epochs": 0,
+ "c_mel": 45,
+ "c_kl": 1.0
+ },
+ "data": {
+ "max_wav_value": 32768.0,
+ "sampling_rate": 40000,
+ "filter_length": 2048,
+ "hop_length": 400,
+ "win_length": 2048,
+ "n_mel_channels": 125,
+ "mel_fmin": 0.0,
+ "mel_fmax": null
+ },
+ "model": {
+ "inter_channels": 192,
+ "hidden_channels": 192,
+ "filter_channels": 768,
+ "n_heads": 2,
+ "n_layers": 6,
+ "kernel_size": 3,
+ "p_dropout": 0,
+ "resblock": "1",
+ "resblock_kernel_sizes": [3,7,11],
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+ "upsample_rates": [10,10,2,2],
+ "upsample_initial_channel": 512,
+ "upsample_kernel_sizes": [16,16,4,4],
+ "use_spectral_norm": false,
+ "gin_channels": 256,
+ "spk_embed_dim": 109
+ }
+}
diff --git a/configs/v1/48k.json b/configs/v1/48k.json
new file mode 100644
index 0000000000000000000000000000000000000000..2d0e05beb794f6f61b769b48c7ae728bf59e6335
--- /dev/null
+++ b/configs/v1/48k.json
@@ -0,0 +1,46 @@
+{
+ "train": {
+ "log_interval": 200,
+ "seed": 1234,
+ "epochs": 20000,
+ "learning_rate": 1e-4,
+ "betas": [0.8, 0.99],
+ "eps": 1e-9,
+ "batch_size": 4,
+ "fp16_run": true,
+ "lr_decay": 0.999875,
+ "segment_size": 11520,
+ "init_lr_ratio": 1,
+ "warmup_epochs": 0,
+ "c_mel": 45,
+ "c_kl": 1.0
+ },
+ "data": {
+ "max_wav_value": 32768.0,
+ "sampling_rate": 48000,
+ "filter_length": 2048,
+ "hop_length": 480,
+ "win_length": 2048,
+ "n_mel_channels": 128,
+ "mel_fmin": 0.0,
+ "mel_fmax": null
+ },
+ "model": {
+ "inter_channels": 192,
+ "hidden_channels": 192,
+ "filter_channels": 768,
+ "n_heads": 2,
+ "n_layers": 6,
+ "kernel_size": 3,
+ "p_dropout": 0,
+ "resblock": "1",
+ "resblock_kernel_sizes": [3,7,11],
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+ "upsample_rates": [10,6,2,2,2],
+ "upsample_initial_channel": 512,
+ "upsample_kernel_sizes": [16,16,4,4,4],
+ "use_spectral_norm": false,
+ "gin_channels": 256,
+ "spk_embed_dim": 109
+ }
+}
diff --git a/configs/v2/32k.json b/configs/v2/32k.json
new file mode 100644
index 0000000000000000000000000000000000000000..70e534f4c641a5a2c8e5c1e172f61398ee97e6e0
--- /dev/null
+++ b/configs/v2/32k.json
@@ -0,0 +1,46 @@
+{
+ "train": {
+ "log_interval": 200,
+ "seed": 1234,
+ "epochs": 20000,
+ "learning_rate": 1e-4,
+ "betas": [0.8, 0.99],
+ "eps": 1e-9,
+ "batch_size": 4,
+ "fp16_run": true,
+ "lr_decay": 0.999875,
+ "segment_size": 12800,
+ "init_lr_ratio": 1,
+ "warmup_epochs": 0,
+ "c_mel": 45,
+ "c_kl": 1.0
+ },
+ "data": {
+ "max_wav_value": 32768.0,
+ "sampling_rate": 32000,
+ "filter_length": 1024,
+ "hop_length": 320,
+ "win_length": 1024,
+ "n_mel_channels": 80,
+ "mel_fmin": 0.0,
+ "mel_fmax": null
+ },
+ "model": {
+ "inter_channels": 192,
+ "hidden_channels": 192,
+ "filter_channels": 768,
+ "n_heads": 2,
+ "n_layers": 6,
+ "kernel_size": 3,
+ "p_dropout": 0,
+ "resblock": "1",
+ "resblock_kernel_sizes": [3,7,11],
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+ "upsample_rates": [10,8,2,2],
+ "upsample_initial_channel": 512,
+ "upsample_kernel_sizes": [20,16,4,4],
+ "use_spectral_norm": false,
+ "gin_channels": 256,
+ "spk_embed_dim": 109
+ }
+}
diff --git a/configs/v2/48k.json b/configs/v2/48k.json
new file mode 100644
index 0000000000000000000000000000000000000000..75f770cdacff3467e9e925ed2393b480881d0303
--- /dev/null
+++ b/configs/v2/48k.json
@@ -0,0 +1,46 @@
+{
+ "train": {
+ "log_interval": 200,
+ "seed": 1234,
+ "epochs": 20000,
+ "learning_rate": 1e-4,
+ "betas": [0.8, 0.99],
+ "eps": 1e-9,
+ "batch_size": 4,
+ "fp16_run": true,
+ "lr_decay": 0.999875,
+ "segment_size": 17280,
+ "init_lr_ratio": 1,
+ "warmup_epochs": 0,
+ "c_mel": 45,
+ "c_kl": 1.0
+ },
+ "data": {
+ "max_wav_value": 32768.0,
+ "sampling_rate": 48000,
+ "filter_length": 2048,
+ "hop_length": 480,
+ "win_length": 2048,
+ "n_mel_channels": 128,
+ "mel_fmin": 0.0,
+ "mel_fmax": null
+ },
+ "model": {
+ "inter_channels": 192,
+ "hidden_channels": 192,
+ "filter_channels": 768,
+ "n_heads": 2,
+ "n_layers": 6,
+ "kernel_size": 3,
+ "p_dropout": 0,
+ "resblock": "1",
+ "resblock_kernel_sizes": [3,7,11],
+ "resblock_dilation_sizes": [[1,3,5], [1,3,5], [1,3,5]],
+ "upsample_rates": [12,10,2,2],
+ "upsample_initial_channel": 512,
+ "upsample_kernel_sizes": [24,20,4,4],
+ "use_spectral_norm": false,
+ "gin_channels": 256,
+ "spk_embed_dim": 109
+ }
+}
diff --git a/docker-compose.yml b/docker-compose.yml
new file mode 100644
index 0000000000000000000000000000000000000000..0768b30333350f9f80afa4870b95e7e472d02256
--- /dev/null
+++ b/docker-compose.yml
@@ -0,0 +1,20 @@
+version: "3.8"
+services:
+ rvc:
+ build:
+ context: .
+ dockerfile: Dockerfile
+ container_name: rvc
+ volumes:
+ - ./weights:/app/assets/weights
+ - ./opt:/app/opt
+ # - ./dataset:/app/dataset # you can use this folder in order to provide your dataset for model training
+ ports:
+ - 7865:7865
+ deploy:
+ resources:
+ reservations:
+ devices:
+ - driver: nvidia
+ count: 1
+ capabilities: [gpu]
\ No newline at end of file
diff --git a/docs/cn/README.cn.md b/docs/cn/README.cn.md
new file mode 100644
index 0000000000000000000000000000000000000000..bf70a4ad53a14087f2fbefa22b1b6a772dd8045d
--- /dev/null
+++ b/docs/cn/README.cn.md
@@ -0,0 +1,198 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+一个基于VITS的简单易用的变声框架
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+[**常见问题解答**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%B8%B8%E8%A7%81%E9%97%AE%E9%A2%98%E8%A7%A3%E7%AD%94) | [**AutoDL·5毛钱训练AI歌手**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B) | [**对照实验记录**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%AF%B9%E7%85%A7%E5%AE%9E%E9%AA%8C%C2%B7%E5%AE%9E%E9%AA%8C%E8%AE%B0%E5%BD%95)) | [**在线演示**](https://modelscope.cn/studios/FlowerCry/RVCv2demo)
+
+[**English**](../../README.md) | [**中文简体**](../cn/README.cn.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+
+
+> 底模使用接近50小时的开源高质量VCTK训练集训练，无版权方面的顾虑，请大家放心使用
+
+> 请期待RVCv3的底模，参数更大，数据集更大，效果更好，基本持平的推理速度，需要训练数据量更少。
+
+> 由于某些地区无法直连Hugging Face，即使设法成功访问，速度也十分缓慢，特推出模型/整合包/工具的一键下载器，欢迎试用：[RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)
+
+| 训练推理界面 |
+| :--------: |
+|  |
+
+| 实时变声界面 |
+| :---------: |
+|  |
+
+## 简介
+本仓库具有以下特点
++ 使用top1检索替换输入源特征为训练集特征来杜绝音色泄漏
++ 即便在相对较差的显卡上也能快速训练
++ 使用少量数据进行训练也能得到较好结果(推荐至少收集10分钟低底噪语音数据)
++ 可以通过模型融合来改变音色(借助ckpt处理选项卡中的ckpt-merge)
++ 简单易用的网页界面
++ 可调用UVR5模型来快速分离人声和伴奏
++ 使用最先进的[人声音高提取算法InterSpeech2023-RMVPE](#参考项目)根绝哑音问题，效果更好，运行更快，资源占用更少
++ A卡I卡加速支持
+
+点此查看我们的[演示视频](https://www.bilibili.com/video/BV1pm4y1z7Gm/) !
+
+## 环境配置
+### Python 版本限制
+> 建议使用 conda 管理 Python 环境
+
+> 版本限制原因参见此[bug](https://github.com/facebookresearch/fairseq/issues/5012)
+
+```bash
+python --version # 3.8 <= Python < 3.11
+```
+
+### Linux/MacOS 一键依赖安装启动脚本
+执行项目根目录下`run.sh`即可一键配置`venv`虚拟环境、自动安装所需依赖并启动主程序。
+```bash
+sh ./run.sh
+```
+
+### 手动安装依赖
+1. 安装`pytorch`及其核心依赖，若已安装则跳过。参考自: https://pytorch.org/get-started/locally/
+ ```bash
+ pip install torch torchvision torchaudio
+ ```
+2. 如果是 win 系统 + Nvidia Ampere 架构(RTX30xx)，根据 #21 的经验，需要指定 pytorch 对应的 CUDA 版本
+ ```bash
+ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+ ```
+3. 根据自己的显卡安装对应依赖
+- N卡
+ ```bash
+ pip install -r requirements/main.txt
+ ```
+- A卡/I卡
+ ```bash
+ pip install -r requirements/dml.txt
+ ```
+- A卡ROCM(Linux)
+ ```bash
+ pip install -r requirements/amd.txt
+ ```
+- I卡IPEX(Linux)
+ ```bash
+ pip install -r requirements/ipex.txt
+ ```
+
+## 其他资源准备
+### 1. assets
+> RVC需要位于`assets`文件夹下的一些模型资源进行推理和训练。
+#### 自动检查/下载资源(默认)
+> 默认情况下，RVC可在主程序启动时自动检查所需资源的完整性。
+
+> 即使资源不完整，程序也将继续启动。
+
+- 如果您希望下载所有资源，请添加`--update`参数
+- 如果您希望跳过启动时的资源完整性检查，请添加`--nocheck`参数
+
+#### 手动下载资源
+> 所有资源文件均位于[Hugging Face space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+> 你可以在`tools`文件夹找到下载它们的脚本
+
+> 你也可以使用模型/整合包/工具的一键下载器：[RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)
+
+以下是一份清单，包括了所有RVC所需的预模型和其他文件的名称。
+
+- ./assets/hubert/hubert_base.pt
+ ```bash
+ rvcmd assets/hubert # RVC-Models-Downloader command
+ ```
+- ./assets/pretrained
+ ```bash
+ rvcmd assets/v1 # RVC-Models-Downloader command
+ ```
+- ./assets/uvr5_weights
+ ```bash
+ rvcmd assets/uvr5 # RVC-Models-Downloader command
+ ```
+想使用v2版本模型的话，需要额外下载
+
+- ./assets/pretrained_v2
+ ```bash
+ rvcmd assets/v2 # RVC-Models-Downloader command
+ ```
+
+### 2. 下载 rmvpe 人声音高提取算法所需文件
+
+如果你想使用最新的RMVPE人声音高提取算法，则你需要下载音高提取模型参数并放置于`assets/rmvpe`。
+
+- 下载[rmvpe.pt](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+#### 下载 rmvpe 的 dml 环境(可选, A卡/I卡用户)
+
+- 下载[rmvpe.onnx](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+### 3. AMD显卡Rocm(可选, 仅Linux)
+
+如果你想基于AMD的Rocm技术在Linux系统上运行RVC，请先在[这里](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html)安装所需的驱动。
+
+若你使用的是Arch Linux，可以使用pacman来安装所需驱动：
+````
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+````
+对于某些型号的显卡，你可能需要额外配置如下的环境变量（如：RX6700XT）：
+````
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+````
+同时确保你的当前用户处于`render`与`video`用户组内：
+````
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+````
+
+## 开始使用
+### 直接启动
+使用以下指令来启动 WebUI
+```bash
+python web.py
+```
+### Linux/MacOS 用户
+```bash
+./run.sh
+```
+### 对于需要使用IPEX技术的I卡用户(仅Linux)
+```bash
+source /opt/intel/oneapi/setvars.sh
+./run.sh
+```
+### 使用整合包 (Windows 用户)
+下载并解压`RVC-beta.7z`，解压后双击`go-web.bat`即可一键启动。
+```bash
+rvcmd packs/general/latest # RVC-Models-Downloader command
+```
+
+## 参考项目
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
++ [Vocal pitch extraction:RMVPE](https://github.com/Dream-High/RMVPE)
+ + The pretrained model is trained and tested by [yxlllc](https://github.com/yxlllc/RMVPE) and [RVC-Boss](https://github.com/RVC-Boss).
+
+## 感谢所有贡献者作出的努力
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
diff --git a/docs/cn/faq.md b/docs/cn/faq.md
new file mode 100644
index 0000000000000000000000000000000000000000..77d9d2aff6de243910512334ce4e7b08cdf632f5
--- /dev/null
+++ b/docs/cn/faq.md
@@ -0,0 +1,150 @@
+## Q1:一键训练结束没有索引
+
+显示"Training is done. The program is closed."则模型训练成功，后续紧邻的报错是假的；
+
+
+一键训练结束完成没有added开头的索引文件，可能是因为训练集太大卡住了添加索引的步骤；已通过批处理add索引解决内存add索引对内存需求过大的问题。临时可尝试再次点击"训练索引"按钮。
+
+
+## Q2:训练结束推理没看到训练集的音色
+点刷新音色再看看，如果还没有看看训练有没有报错，控制台和webui的截图，logs/实验名下的log，都可以发给开发者看看。
+
+
+## Q3:如何分享模型
+ rvc_root/logs/实验名 下面存储的pth不是用来分享模型用来推理的，而是为了存储实验状态供复现，以及继续训练用的。用来分享的模型应该是weights文件夹下大小为60+MB的pth文件；
+
+ 后续将把weights/exp_name.pth和logs/exp_name/added_xxx.index合并打包成weights/exp_name.zip省去填写index的步骤，那么zip文件用来分享，不要分享pth文件，除非是想换机器继续训练；
+
+ 如果你把logs文件夹下的几百MB的pth文件复制/分享到weights文件夹下强行用于推理，可能会出现f0，tgt_sr等各种key不存在的报错。你需要用ckpt选项卡最下面，手工或自动（本地logs下如果能找到相关信息则会自动）选择是否携带音高、目标音频采样率的选项后进行ckpt小模型提取（输入路径填G开头的那个），提取完在weights文件夹下会出现60+MB的pth文件，刷新音色后可以选择使用。
+
+
+## Q4:Connection Error.
+也许你关闭了控制台（黑色窗口）。
+
+
+## Q5:WebUI弹出Expecting value: line 1 column 1 (char 0).
+请关闭系统局域网代理/全局代理。
+
+
+这个不仅是客户端的代理，也包括服务端的代理（例如你使用autodl设置了http_proxy和https_proxy学术加速，使用时也需要unset关掉）
+
+
+## Q6:不用WebUI如何通过命令训练推理
+训练脚本：
+
+可先跑通WebUI，消息窗内会显示数据集处理和训练用命令行；
+
+
+推理脚本：
+
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+
+例子：
+
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+
+f0up_key=sys.argv[1]
+
+input_path=sys.argv[2]
+
+index_path=sys.argv[3]
+
+f0method=sys.argv[4]#harvest or pm
+
+opt_path=sys.argv[5]
+
+model_path=sys.argv[6]
+
+index_rate=float(sys.argv[7])
+
+device=sys.argv[8]
+
+is_half=bool(sys.argv[9])
+
+
+## Q7:Cuda error/Cuda out of memory.
+小概率是cuda配置问题、设备不支持；大概率是显存不够（out of memory）；
+
+
+训练的话缩小batch size（如果缩小到1还不够只能更换显卡训练），推理的话酌情缩小config.py结尾的x_pad，x_query，x_center，x_max。4G以下显存（例如1060（3G）和各种2G显卡）可以直接放弃，4G显存显卡还有救。
+
+
+## Q8:total_epoch调多少比较好
+
+如果训练集音质差底噪大，20~30足够了，调太高，底模音质无法带高你的低音质训练集
+
+如果训练集音质高底噪低时长多，可以调高，200是ok的（训练速度很快，既然你有条件准备高音质训练集，显卡想必条件也不错，肯定不在乎多一些训练时间）
+
+
+## Q9:需要多少训练集时长
+ 推荐10min至50min
+
+ 保证音质高底噪低的情况下，如果有个人特色的音色统一，则多多益善
+
+ 高水平的训练集（精简+音色有特色），5min至10min也是ok的，仓库作者本人就经常这么玩
+
+ 也有人拿1min至2min的数据来训练并且训练成功的，但是成功经验是其他人不可复现的，不太具备参考价值。这要求训练集音色特色非常明显（比如说高频气声较明显的萝莉少女音），且音质高；
+
+ 1min以下时长数据目前没见有人尝试（成功）过。不建议进行这种鬼畜行为。
+
+
+## Q10:index rate干嘛用的，怎么调（科普）
+ 如果底模和推理源的音质高于训练集的音质，他们可以带高推理结果的音质，但代价可能是音色往底模/推理源的音色靠，这种现象叫做"音色泄露"；
+
+ index rate用来削减/解决音色泄露问题。调到1，则理论上不存在推理源的音色泄露问题，但音质更倾向于训练集。如果训练集音质比推理源低，则index rate调高可能降低音质。调到0，则不具备利用检索混合来保护训练集音色的效果；
+
+ 如果训练集优质时长多，可调高total_epoch，此时模型本身不太会引用推理源和底模的音色，很少存在"音色泄露"问题，此时index_rate不重要，你甚至可以不建立/分享index索引文件。
+
+
+## Q11:推理怎么选gpu
+config.py文件里device cuda:后面选择卡号；
+
+卡号和显卡的映射关系，在训练选项卡的显卡信息栏里能看到。
+
+
+## Q12:如何推理训练中间保存的pth
+通过ckpt选项卡最下面提取小模型。
+
+
+
+## Q13:如何中断和继续训练
+现阶段只能关闭WebUI控制台双击go-web.bat重启程序。网页参数也要刷新重新填写；
+
+继续训练：相同网页参数点训练模型，就会接着上次的checkpoint继续训练。
+
+
+## Q14:训练时出现文件页面/内存error
+进程开太多了，内存炸了。你可能可以通过如下方式解决
+
+1、"提取音高和处理数据使用的CPU进程数" 酌情拉低；
+
+2、训练集音频手工切一下，不要太长。
+
+
+
+## Q15:如何中途加数据训练
+1、所有数据新建一个实验名；
+
+2、拷贝上一次的最新的那个G和D文件（或者你想基于哪个中间ckpt训练，也可以拷贝中间的）到新实验名；下
+
+3、一键训练新实验名，他会继续上一次的最新进度训练。
+
+
+## Q16: error about llvmlite.dll
+
+OSError: Could not load shared object file: llvmlite.dll
+
+FileNotFoundError: Could not find module lib\site-packages\llvmlite\binding\llvmlite.dll (or one of its dependencies). Try using the full path with constructor syntax.
+
+win平台会报这个错，装上https://aka.ms/vs/17/release/vc_redist.x64.exe这个再重启WebUI就好了。
+
+## Q17: RuntimeError: The expanded size of the tensor (17280) must match the existing size (0) at non-singleton dimension 1. Target sizes: [1, 17280]. Tensor sizes: [0]
+
+wavs16k文件夹下，找到文件大小显著比其他都小的一些音频文件，删掉，点击训练模型，就不会报错了，不过由于一键流程中断了你训练完模型还要点训练索引。
+
+## Q18: RuntimeError: The size of tensor a (24) must match the size of tensor b (16) at non-singleton dimension 2
+
+不要中途变更采样率继续训练。如果一定要变更，应更换实验名从头训练。当然你也可以把上次提取的音高和特征（0/1/2/2b folders）拷贝过去加速训练流程。
diff --git a/docs/en/faiss_tips_en.md b/docs/en/faiss_tips_en.md
new file mode 100644
index 0000000000000000000000000000000000000000..aafad6ed67f70ee1ea3a2a21ee0b5066ab1dcfa8
--- /dev/null
+++ b/docs/en/faiss_tips_en.md
@@ -0,0 +1,102 @@
+faiss tuning TIPS
+==================
+# about faiss
+faiss is a library of neighborhood searches for dense vectors, developed by facebook research, which efficiently implements many approximate neighborhood search methods.
+Approximate Neighbor Search finds similar vectors quickly while sacrificing some accuracy.
+
+## faiss in RVC
+In RVC, for the embedding of features converted by HuBERT, we search for embeddings similar to the embedding generated from the training data and mix them to achieve a conversion that is closer to the original speech. However, since this search takes time if performed naively, high-speed conversion is realized by using approximate neighborhood search.
+
+# implementation overview
+In '/logs/your-experiment/3_feature256' where the model is located, features extracted by HuBERT from each voice data are located.
+From here we read the npy files in order sorted by filename and concatenate the vectors to create big_npy. (This vector has shape [N, 256].)
+After saving big_npy as /logs/your-experiment/total_fea.npy, train it with faiss.
+
+In this article, I will explain the meaning of these parameters.
+
+# Explanation of the method
+## index factory
+An index factory is a unique faiss notation that expresses a pipeline that connects multiple approximate neighborhood search methods as a string.
+This allows you to try various approximate neighborhood search methods simply by changing the index factory string.
+In RVC it is used like this:
+
+```python
+index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+Among the arguments of index_factory, the first is the number of dimensions of the vector, the second is the index factory string, and the third is the distance to use.
+
+For more detailed notation
+https://github.com/facebookresearch/faiss/wiki/The-index-factory
+
+## index for distance
+There are two typical indexes used as similarity of embedding as follows.
+
+- Euclidean distance (METRIC_L2)
+- inner product (METRIC_INNER_PRODUCT)
+
+Euclidean distance takes the squared difference in each dimension, sums the differences in all dimensions, and then takes the square root. This is the same as the distance in 2D and 3D that we use on a daily basis.
+The inner product is not used as an index of similarity as it is, and the cosine similarity that takes the inner product after being normalized by the L2 norm is generally used.
+
+Which is better depends on the case, but cosine similarity is often used in embedding obtained by word2vec and similar image retrieval models learned by ArcFace. If you want to do l2 normalization on vector X with numpy, you can do it with the following code with eps small enough to avoid 0 division.
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+Also, for the index factory, you can change the distance index used for calculation by choosing the value to pass as the third argument.
+
+```python
+index = faiss.index_factory(dimention, text, faiss.METRIC_INNER_PRODUCT)
+```
+
+## IVF
+IVF (Inverted file indexes) is an algorithm similar to the inverted index in full-text search.
+During learning, the search target is clustered with kmeans, and Voronoi partitioning is performed using the cluster center. Each data point is assigned a cluster, so we create a dictionary that looks up the data points from the clusters.
+
+For example, if clusters are assigned as follows
+|index|Cluster|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+The resulting inverted index looks like this:
+
+|cluster|index|
+|-------|-----|
+|A|1, 3|
+|B|2, 5|
+|C|4|
+
+When searching, we first search n_probe clusters from the clusters, and then calculate the distances for the data points belonging to each cluster.
+
+# recommend parameter
+There are official guidelines on how to choose an index, so I will explain accordingly.
+https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
+
+For datasets below 1M, 4bit-PQ is the most efficient method available in faiss as of April 2023.
+Combining this with IVF, narrowing down the candidates with 4bit-PQ, and finally recalculating the distance with an accurate index can be described by using the following index factory.
+
+```python
+index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## Recommended parameters for IVF
+Consider the case of too many IVFs. For example, if coarse quantization by IVF is performed for the number of data, this is the same as a naive exhaustive search and is inefficient.
+For 1M or less, IVF values are recommended between 4*sqrt(N) ~ 16*sqrt(N) for N number of data points.
+
+Since the calculation time increases in proportion to the number of n_probes, please consult with the accuracy and choose appropriately. Personally, I don't think RVC needs that much accuracy, so n_probe = 1 is fine.
+
+## FastScan
+FastScan is a method that enables high-speed approximation of distances by Cartesian product quantization by performing them in registers.
+Cartesian product quantization performs clustering independently for each d dimension (usually d = 2) during learning, calculates the distance between clusters in advance, and creates a lookup table. At the time of prediction, the distance of each dimension can be calculated in O(1) by looking at the lookup table.
+So the number you specify after PQ usually specifies half the dimension of the vector.
+
+For a more detailed description of FastScan, please refer to the official documentation.
+https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlat is an instruction to recalculate the rough distance calculated by FastScan with the exact distance specified by the third argument of index factory.
+When getting k neighbors, k*k_factor points are recalculated.
diff --git a/docs/en/faq_en.md b/docs/en/faq_en.md
new file mode 100644
index 0000000000000000000000000000000000000000..23e325c2474ffce7fb998c32e99fc6139a55757b
--- /dev/null
+++ b/docs/en/faq_en.md
@@ -0,0 +1,114 @@
+## Q1:Cannot find index file after "One-click Training".
+If it displays "Training is done. The program is closed," then the model has been trained successfully, and the subsequent errors are fake;
+
+The lack of an 'added' index file after One-click training may be due to the training set being too large, causing the addition of the index to get stuck; this has been resolved by using batch processing to add the index, which solves the problem of memory overload when adding the index. As a temporary solution, try clicking the "Train Index" button again.
+
+## Q2:Cannot find the model in “Inferencing timbre” after training
+Click “Refresh timbre list” and check again; if still not visible, check if there are any errors during training and send screenshots of the console, web UI, and logs/experiment_name/*.log to the developers for further analysis.
+
+## Q3:How to share a model/How to use others' models?
+The pth files stored in rvc_root/logs/experiment_name are not meant for sharing or inference, but for storing the experiment checkpoits for reproducibility and further training. The model to be shared should be the 60+MB pth file in the weights folder;
+
+In the future, weights/exp_name.pth and logs/exp_name/added_xxx.index will be merged into a single weights/exp_name.zip file to eliminate the need for manual index input; so share the zip file, not the pth file, unless you want to continue training on a different machine;
+
+Copying/sharing the several hundred MB pth files from the logs folder to the weights folder for forced inference may result in errors such as missing f0, tgt_sr, or other keys. You need to use the ckpt tab at the bottom to manually or automatically (if the information is found in the logs/exp_name), select whether to include pitch infomation and target audio sampling rate options and then extract the smaller model. After extraction, there will be a 60+ MB pth file in the weights folder, and you can refresh the voices to use it.
+
+## Q4:Connection Error.
+You may have closed the console (black command line window).
+
+## Q5:WebUI popup 'Expecting value: line 1 column 1 (char 0)'.
+Please disable system LAN proxy/global proxy and then refresh.
+
+## Q6:How to train and infer without the WebUI?
+Training script:
+You can run training in WebUI first, and the command-line versions of dataset preprocessing and training will be displayed in the message window.
+
+Inference script:
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+
+e.g.
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#harvest or pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+## Q7:Cuda error/Cuda out of memory.
+There is a small chance that there is a problem with the CUDA configuration or the device is not supported; more likely, there is not enough memory (out of memory).
+
+For training, reduce the batch size (if reducing to 1 is still not enough, you may need to change the graphics card); for inference, adjust the x_pad, x_query, x_center, and x_max settings in the config.py file as needed. 4G or lower memory cards (e.g. 1060(3G) and various 2G cards) can be abandoned, while 4G memory cards still have a chance.
+
+## Q8:How many total_epoch are optimal?
+If the training dataset's audio quality is poor and the noise floor is high, 20-30 epochs are sufficient. Setting it too high won't improve the audio quality of your low-quality training set.
+
+If the training set audio quality is high, the noise floor is low, and there is sufficient duration, you can increase it. 200 is acceptable (since training is fast, and if you're able to prepare a high-quality training set, your GPU likely can handle a longer training duration without issue).
+
+## Q9:How much training set duration is needed?
+
+A dataset of around 10min to 50min is recommended.
+
+With guaranteed high sound quality and low bottom noise, more can be added if the dataset's timbre is uniform.
+
+For a high-level training set (lean + distinctive tone), 5min to 10min is fine.
+
+There are some people who have trained successfully with 1min to 2min data, but the success is not reproducible by others and is not very informative.
This requires that the training set has a very distinctive timbre (e.g. a high-frequency airy anime girl sound) and the quality of the audio is high;
+Data of less than 1min duration has not been successfully attempted so far. This is not recommended.
+
+
+## Q10:What is the index rate for and how to adjust it?
+If the tone quality of the pre-trained model and inference source is higher than that of the training set, they can bring up the tone quality of the inference result, but at the cost of a possible tone bias towards the tone of the underlying model/inference source rather than the tone of the training set, which is generally referred to as "tone leakage".
+
+The index rate is used to reduce/resolve the timbre leakage problem. If the index rate is set to 1, theoretically there is no timbre leakage from the inference source and the timbre quality is more biased towards the training set. If the training set has a lower sound quality than the inference source, then a higher index rate may reduce the sound quality. Turning it down to 0 does not have the effect of using retrieval blending to protect the training set tones.
+
+If the training set has good audio quality and long duration, turn up the total_epoch, when the model itself is less likely to refer to the inferred source and the pretrained underlying model, and there is little "tone leakage", the index_rate is not important and you can even not create/share the index file.
+
+## Q11:How to choose the gpu when inferring?
+In the config.py file, select the card number after "device cuda:".
+
+The mapping between card number and graphics card can be seen in the graphics card information section of the training tab.
+
+## Q12:How to use the model saved in the middle of training?
+Save via model extraction at the bottom of the ckpt processing tab.
+
+## Q13:File/memory error(when training)?
+Too many processes and your memory is not enough. You may fix it by:
+
+1、decrease the input in field "Threads of CPU".
+
+2、pre-cut trainset to shorter audio files.
+
+## Q14: How to continue training using more data
+
+step1: put all wav data to path2.
+
+step2: exp_name2+path2 -> process dataset and extract feature.
+
+step3: copy the latest G and D file of exp_name1 (your previous experiment) into exp_name2 folder.
+
+step4: click "train the model", and it will continue training from the beginning of your previous exp model epoch.
+
+## Q15: error about llvmlite.dll
+
+OSError: Could not load shared object file: llvmlite.dll
+
+FileNotFoundError: Could not find module lib\site-packages\llvmlite\binding\llvmlite.dll (or one of its dependencies). Try using the full path with constructor syntax.
+
+The issue will happen in windows, install https://aka.ms/vs/17/release/vc_redist.x64.exe and it will be fixed.
+
+## Q16: RuntimeError: The expanded size of the tensor (17280) must match the existing size (0) at non-singleton dimension 1. Target sizes: [1, 17280]. Tensor sizes: [0]
+
+Delete the wav files whose size is significantly smaller than others, and that won't happen again. Than click "train the model"and "train the index".
+
+## Q17: RuntimeError: The size of tensor a (24) must match the size of tensor b (16) at non-singleton dimension 2
+
+Do not change the sampling rate and then continue training. If it is necessary to change, the exp name should be changed and the model will be trained from scratch. You can also copy the pitch and features (0/1/2/2b folders) extracted last time to accelerate the training process.
+
diff --git a/docs/en/training_tips_en.md b/docs/en/training_tips_en.md
new file mode 100644
index 0000000000000000000000000000000000000000..2aca8a77a864298a890b8b926505a6eef5f27319
--- /dev/null
+++ b/docs/en/training_tips_en.md
@@ -0,0 +1,65 @@
+Instructions and tips for RVC training
+======================================
+This TIPS explains how data training is done.
+
+# Training flow
+I will explain along the steps in the training tab of the GUI.
+
+## step1
+Set the experiment name here.
+
+You can also set here whether the model should take pitch into account.
+If the model doesn't consider pitch, the model will be lighter, but not suitable for singing.
+
+Data for each experiment is placed in `/logs/your-experiment-name/`.
+
+## step2a
+Loads and preprocesses audio.
+
+### load audio
+If you specify a folder with audio, the audio files in that folder will be read automatically.
+For example, if you specify `C:Users\hoge\voices`, `C:Users\hoge\voices\voice.mp3` will be loaded, but `C:Users\hoge\voices\dir\voice.mp3` will Not loaded.
+
+Since ffmpeg is used internally for reading audio, if the extension is supported by ffmpeg, it will be read automatically.
+After converting to int16 with ffmpeg, convert to float32 and normalize between -1 to 1.
+
+### denoising
+The audio is smoothed by scipy's filtfilt.
+
+### Audio Split
+First, the input audio is divided by detecting parts of silence that last longer than a certain period (max_sil_kept=5 seconds?). After splitting the audio on silence, split the audio every 4 seconds with an overlap of 0.3 seconds. For audio separated within 4 seconds, after normalizing the volume, convert the wav file to `/logs/your-experiment-name/0_gt_wavs` and then convert it to 16k sampling rate to `/logs/your-experiment-name/1_16k_wavs ` as a wav file.
+
+## step2b
+### Extract pitch
+Extract pitch information from wav files. Extract the pitch information (=f0) using the method built into parselmouth or pyworld and save it in `/logs/your-experiment-name/2a_f0`. Then logarithmically convert the pitch information to an integer between 1 and 255 and save it in `/logs/your-experiment-name/2b-f0nsf`.
+
+### Extract feature_print
+Convert the wav file to embedding in advance using HuBERT. Read the wav file saved in `/logs/your-experiment-name/1_16k_wavs`, convert the wav file to 256-dimensional features with HuBERT, and save in npy format in `/logs/your-experiment-name/3_feature256`.
+
+## step3
+train the model.
+### Glossary for Beginners
+In deep learning, the data set is divided and the learning proceeds little by little. In one model update (step), batch_size data are retrieved and predictions and error corrections are performed. Doing this once for a dataset counts as one epoch.
+
+Therefore, the learning time is the learning time per step x (the number of data in the dataset / batch size) x the number of epochs. In general, the larger the batch size, the more stable the learning becomes (learning time per step ÷ batch size) becomes smaller, but it uses more GPU memory. GPU RAM can be checked with the nvidia-smi command. Learning can be done in a short time by increasing the batch size as much as possible according to the machine of the execution environment.
+
+### Specify pretrained model
+RVC starts training the model from pretrained weights instead of from 0, so it can be trained with a small dataset.
+
+By default
+
+- If you consider pitch, it loads `rvc-location/pretrained/f0G40k.pth` and `rvc-location/pretrained/f0D40k.pth`.
+- If you don't consider pitch, it loads `rvc-location/pretrained/G40k.pth` and `rvc-location/pretrained/D40k.pth`.
+
+When learning, model parameters are saved in `logs/your-experiment-name/G_{}.pth` and `logs/your-experiment-name/D_{}.pth` for each save_every_epoch, but by specifying this path, you can start learning. You can restart or start training from model weights learned in a different experiment.
+
+### learning index
+RVC saves the HuBERT feature values used during training, and during inference, searches for feature values that are similar to the feature values used during learning to perform inference. In order to perform this search at high speed, the index is learned in advance.
+For index learning, we use the approximate neighborhood search library faiss. Read the feature value of `logs/your-experiment-name/3_feature256` and use it to learn the index, and save it as `logs/your-experiment-name/add_XXX.index`.
+
+(From the 20230428update version, it is read from the index, and saving / specifying is no longer necessary.)
+
+### Button description
+- Train model: After executing step2b, press this button to train the model.
+- Train feature index: After training the model, perform index learning.
+- One-click training: step2b, model training and feature index training all at once.
\ No newline at end of file
diff --git a/docs/fr/README.fr.md b/docs/fr/README.fr.md
new file mode 100644
index 0000000000000000000000000000000000000000..18699232a30bd47dbb3ca2d8d5c15ba62ae537b6
--- /dev/null
+++ b/docs/fr/README.fr.md
@@ -0,0 +1,166 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+Un framework simple et facile à utiliser pour la conversion vocale (modificateur de voix) basé sur VITS
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+[**FAQ**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%B8%B8%E8%A7%81%E9%97%AE%E9%A2%98%E8%A7%A3%E7%AD%94) | [**AutoDL·Formation d'un chanteur AI pour 5 centimes**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B) | [**Enregistrement des expériences comparatives**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%AF%B9%E7%85%A7%E5%AE%9E%E9%AA%8C%C2%B7%E5%AE%9E%E9%AA%8C%E8%AE%B0%E5%BD%95) | [**Démonstration en ligne**](https://huggingface.co/spaces/Ricecake123/RVC-demo)
+
+
+
+------
+
+[**English**](../en/README.en.md) | [ **中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Turc**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+Cliquez ici pour voir notre [vidéo de démonstration](https://www.bilibili.com/video/BV1pm4y1z7Gm/) !
+
+> Conversion vocale en temps réel avec RVC : [w-okada/voice-changer](https://github.com/w-okada/voice-changer)
+
+> Le modèle de base est formé avec près de 50 heures de données VCTK de haute qualité et open source. Aucun souci concernant les droits d'auteur, n'hésitez pas à l'utiliser.
+
+> Attendez-vous au modèle de base RVCv3 : plus de paramètres, plus de données, de meilleurs résultats, une vitesse d'inférence presque identique, et nécessite moins de données pour la formation.
+
+## Introduction
+Ce dépôt a les caractéristiques suivantes :
++ Utilise le top1 pour remplacer les caractéristiques de la source d'entrée par les caractéristiques de l'ensemble d'entraînement pour éliminer les fuites de timbre vocal.
++ Peut être formé rapidement même sur une carte graphique relativement moins performante.
++ Obtient de bons résultats même avec peu de données pour la formation (il est recommandé de collecter au moins 10 minutes de données vocales avec un faible bruit de fond).
++ Peut changer le timbre vocal en fusionnant des modèles (avec l'aide de l'onglet ckpt-merge).
++ Interface web simple et facile à utiliser.
++ Peut appeler le modèle UVR5 pour séparer rapidement la voix et l'accompagnement.
++ Utilise l'algorithme de pitch vocal le plus avancé [InterSpeech2023-RMVPE](#projets-référencés) pour éliminer les problèmes de voix muette. Meilleurs résultats, plus rapide que crepe_full, et moins gourmand en ressources.
++ Support d'accélération pour les cartes AMD et Intel.
+
+## Configuration de l'environnement
+Exécutez les commandes suivantes dans un environnement Python de version 3.8 ou supérieure.
+
+(Windows/Linux)
+Installez d'abord les dépendances principales via pip :
+```bash
+# Installez Pytorch et ses dépendances essentielles, sautez si déjà installé.
+# Voir : https://pytorch.org/get-started/locally/
+pip install torch torchvision torchaudio
+
+# Pour les utilisateurs de Windows avec une architecture Nvidia Ampere (RTX30xx), en se basant sur l'expérience #21, spécifiez la version CUDA correspondante pour Pytorch.
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+
+# Pour Linux + carte AMD, utilisez cette version de Pytorch:
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
+```
+
+Vous pouvez utiliser poetry pour installer les dépendances :
+```bash
+# Installez l'outil de gestion des dépendances Poetry, sautez si déjà installé.
+# Voir : https://python-poetry.org/docs/#installation
+curl -sSL https://install.python-poetry.org | python3 -
+
+# Installez les dépendances avec poetry.
+poetry install
+```
+
+Ou vous pouvez utiliser pip pour installer les dépendances :
+```bash
+# Cartes Nvidia :
+pip install -r requirements/main.txt
+
+# Cartes AMD/Intel :
+pip install -r requirements/dml.txt
+
+# Cartes Intel avec IPEX
+pip install -r requirements/ipex.txt
+
+# Cartes AMD sur Linux (ROCm)
+pip install -r requirements/amd.txt
+```
+
+------
+Les utilisateurs de Mac peuvent exécuter `run.sh` pour installer les dépendances :
+```bash
+sh ./run.sh
+```
+
+## Préparation d'autres modèles pré-entraînés
+RVC nécessite d'autres modèles pré-entraînés pour l'inférence et la formation.
+
+```bash
+#Télécharger tous les modèles depuis https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/
+python tools/download_models.py
+```
+
+Ou vous pouvez télécharger ces modèles depuis notre [espace Hugging Face](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/).
+
+Voici une liste des modèles et autres fichiers requis par RVC :
+```bash
+./assets/hubert/hubert_base.pt
+
+./assets/pretrained
+
+./assets/uvr5_weights
+
+# Pour tester la version v2 du modèle, téléchargez également :
+
+./assets/pretrained_v2
+
+# Si vous souhaitez utiliser le dernier algorithme RMVPE de pitch vocal, téléchargez les paramètres du modèle de pitch et placez-les dans le répertoire racine de RVC.
+
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt
+
+ # Les utilisateurs de cartes AMD/Intel nécessitant l'environnement DML doivent télécharger :
+
+ https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx
+
+```
+Pour les utilisateurs d'Intel ARC avec IPEX, exécutez d'abord `source /opt/intel/oneapi/setvars.sh`.
+Ensuite, exécutez la commande suivante pour démarrer WebUI :
+```bash
+python web.py
+```
+
+Si vous utilisez Windows ou macOS, vous pouvez télécharger et extraire `RVC-beta.7z`. Les utilisateurs de Windows peuvent exécuter `go-web.bat` pour démarrer WebUI, tandis que les utilisateurs de macOS peuvent exécuter `sh ./run.sh`.
+
+## Compatibilité ROCm pour les cartes AMD (seulement Linux)
+Installez tous les pilotes décrits [ici](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html).
+
+Sur Arch utilisez pacman pour installer le pilote:
+````
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+````
+
+Vous devrez peut-être créer ces variables d'environnement (par exemple avec RX6700XT):
+````
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+````
+Assurez-vous que votre utilisateur est dans les groupes `render` et `video`:
+````
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+````
+Enfin vous pouvez exécuter WebUI:
+```bash
+python web.py
+```
+
+## Crédits
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
++ [Extraction de la hauteur vocale : RMVPE](https://github.com/Dream-High/RMVPE)
+ + Le modèle pré-entraîné a été formé et testé par [yxlllc](https://github.com/yxlllc/RMVPE) et [RVC-Boss](https://github.com/RVC-Boss).
+
+## Remerciements à tous les contributeurs pour leurs efforts
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
diff --git a/docs/fr/faiss_tips_fr.md b/docs/fr/faiss_tips_fr.md
new file mode 100644
index 0000000000000000000000000000000000000000..7fde76a46022e9a5f6ae5fd7315176dabbd375d4
--- /dev/null
+++ b/docs/fr/faiss_tips_fr.md
@@ -0,0 +1,105 @@
+Conseils de réglage pour faiss
+==================
+# À propos de faiss
+faiss est une bibliothèque de recherches de voisins pour les vecteurs denses, développée par Facebook Research, qui implémente efficacement de nombreuses méthodes de recherche de voisins approximatifs.
+La recherche de voisins approximatifs trouve rapidement des vecteurs similaires tout en sacrifiant une certaine précision.
+
+## faiss dans RVC
+Dans RVC, pour l'incorporation des caractéristiques converties par HuBERT, nous recherchons des incorporations similaires à l'incorporation générée à partir des données d'entraînement et les mixons pour obtenir une conversion plus proche de la parole originale. Cependant, cette recherche serait longue si elle était effectuée de manière naïve, donc une conversion à haute vitesse est réalisée en utilisant une recherche de voisinage approximatif.
+
+# Vue d'ensemble de la mise en œuvre
+Dans '/logs/votre-expérience/3_feature256' où le modèle est situé, les caractéristiques extraites par HuBERT de chaque donnée vocale sont situées.
+À partir de là, nous lisons les fichiers npy dans un ordre trié par nom de fichier et concaténons les vecteurs pour créer big_npy. (Ce vecteur a la forme [N, 256].)
+Après avoir sauvegardé big_npy comme /logs/votre-expérience/total_fea.npy, nous l'entraînons avec faiss.
+
+Dans cet article, j'expliquerai la signification de ces paramètres.
+
+# Explication de la méthode
+## Usine d'index
+Une usine d'index est une notation unique de faiss qui exprime un pipeline qui relie plusieurs méthodes de recherche de voisinage approximatif sous forme de chaîne.
+Cela vous permet d'essayer diverses méthodes de recherche de voisinage approximatif simplement en changeant la chaîne de l'usine d'index.
+Dans RVC, elle est utilisée comme ceci :
+
+```python
+index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+
+Parmi les arguments de index_factory, le premier est le nombre de dimensions du vecteur, le second est la chaîne de l'usine d'index, et le troisième est la distance à utiliser.
+
+Pour une notation plus détaillée :
+https://github.com/facebookresearch/faiss/wiki/The-index-factory
+
+## Index pour la distance
+Il existe deux index typiques utilisés comme similarité de l'incorporation comme suit :
+
+- Distance euclidienne (METRIC_L2)
+- Produit intérieur (METRIC_INNER_PRODUCT)
+
+La distance euclidienne prend la différence au carré dans chaque dimension, somme les différences dans toutes les dimensions, puis prend la racine carrée. C'est la même chose que la distance en 2D et 3D que nous utilisons au quotidien.
+Le produit intérieur n'est pas utilisé comme index de similarité tel quel, et la similarité cosinus qui prend le produit intérieur après avoir été normalisé par la norme L2 est généralement utilisée.
+
+Lequel est le mieux dépend du cas, mais la similarité cosinus est souvent utilisée dans l'incorporation obtenue par word2vec et des modèles de récupération d'images similaires appris par ArcFace. Si vous voulez faire une normalisation l2 sur le vecteur X avec numpy, vous pouvez le faire avec le code suivant avec eps suffisamment petit pour éviter une division par 0.
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+De plus, pour l'usine d'index, vous pouvez changer l'index de distance utilisé pour le calcul en choisissant la valeur à passer comme troisième argument.
+
+```python
+index = faiss.index_factory(dimention, texte, faiss.METRIC_INNER_PRODUCT)
+```
+
+## IVF
+IVF (Inverted file indexes) est un algorithme similaire à l'index inversé dans la recherche en texte intégral.
+Lors de l'apprentissage, la cible de recherche est regroupée avec kmeans, et une partition de Voronoi est effectuée en utilisant le centre du cluster. Chaque point de données est attribué à un cluster, nous créons donc un dictionnaire qui permet de rechercher les points de données à partir des clusters.
+
+Par exemple, si des clusters sont attribués comme suit :
+|index|Cluster|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+L'index inversé résultant ressemble à ceci :
+
+|cluster|index|
+|-------|-----|
+|A|1, 3|
+|B|2, 5|
+|C|4|
+
+Lors de la recherche, nous recherchons d'abord n_probe clusters parmi les clusters, puis nous calculons les distances pour les points de données appartenant à chaque cluster.
+
+# Recommandation de paramètre
+Il existe des directives officielles sur la façon de choisir un index, je vais donc expliquer en conséquence.
+https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
+
+Pour les ensembles de données inférieurs à 1M, 4bit-PQ est la méthode la plus efficace disponible dans faiss en avril 2023.
+En combinant cela avec IVF, en réduisant les candidats avec 4bit-PQ, et enfin en recalculant la distance avec un index précis, on peut le décrire en utilisant l'usine d'index suivante.
+
+```python
+index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## Paramètres recommandés pour IVF
+Considérez le cas de trop d'IVF. Par exemple, si une quantification grossière par IVF est effectuée pour le nombre de données, cela revient à une recherche exhaustive naïve et est inefficace.
+Pour 1M ou moins, les valeurs IVF sont recommandées entre 4*sqrt(N) ~ 16*sqrt(N) pour N nombre de points de données.
+
+Comme le temps de calcul augmente proportionnellement au nombre de n_probes, veuillez consulter la précision et choisir de manière appropriée. Personnellement, je ne pense pas que RVC ait besoin de tant de précision, donc n_probe = 1 est bien.
+
+## FastScan
+FastScan est une méthode qui permet d'approximer rapidement les distances par quantification de produit cartésien en les effectuant dans les registres.
+La quantification du produit cartésien effectue un regroupement indépendamment
+
+ pour chaque dimension d (généralement d = 2) pendant l'apprentissage, calcule la distance entre les clusters à l'avance, et crée une table de recherche. Au moment de la prédiction, la distance de chaque dimension peut être calculée en O(1) en consultant la table de recherche.
+Le nombre que vous spécifiez après PQ spécifie généralement la moitié de la dimension du vecteur.
+
+Pour une description plus détaillée de FastScan, veuillez consulter la documentation officielle.
+https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlat est une instruction pour recalculer la distance approximative calculée par FastScan avec la distance exacte spécifiée par le troisième argument de l'usine d'index.
+Lors de l'obtention de k voisins, k*k_factor points sont recalculés.
diff --git a/docs/fr/faq_fr.md b/docs/fr/faq_fr.md
new file mode 100644
index 0000000000000000000000000000000000000000..e0b071f806364beccc9580dc40ad5ffb471256d8
--- /dev/null
+++ b/docs/fr/faq_fr.md
@@ -0,0 +1,164 @@
+## Q1: Impossible de trouver le fichier index après "Entraînement en un clic".
+Si l'affichage indique "L'entraînement est terminé. Le programme est fermé", alors le modèle a été formé avec succès, et les erreurs subséquentes sont fausses ;
+
+L'absence d'un fichier index 'ajouté' après un entraînement en un clic peut être due au fait que le jeu d'entraînement est trop grand, ce qui bloque l'ajout de l'index ; cela a été résolu en utilisant un traitement par lots pour ajouter l'index, ce qui résout le problème de surcharge de mémoire lors de l'ajout de l'index. Comme solution temporaire, essayez de cliquer à nouveau sur le bouton "Entraîner l'index".
+
+## Q2: Impossible de trouver le modèle dans “Inférence du timbre” après l'entraînement
+Cliquez sur “Actualiser la liste des timbres” et vérifiez à nouveau ; si vous ne le voyez toujours pas, vérifiez s'il y a des erreurs pendant l'entraînement et envoyez des captures d'écran de la console, de l'interface utilisateur web, et des logs/nom_de_l'expérience/*.log aux développeurs pour une analyse plus approfondie.
+
+## Q3: Comment partager un modèle/Comment utiliser les modèles d'autres personnes ?
+Les fichiers pth stockés dans rvc_root/logs/nom_de_l'expérience ne sont pas destinés à être partagés ou inférés, mais à stocker les points de contrôle de l'expérience pour la reproductibilité et l'entraînement ultérieur. Le modèle à partager doit être le fichier pth de 60+MB dans le dossier des poids ;
+
+À l'avenir, les poids/nom_de_l'expérience.pth et les logs/nom_de_l'expérience/ajouté_xxx.index seront fusionnés en un seul fichier poids/nom_de_l'expérience.zip pour éliminer le besoin d'une entrée d'index manuelle ; partagez donc le fichier zip, et non le fichier pth, sauf si vous souhaitez continuer l'entraînement sur une machine différente ;
+
+Copier/partager les fichiers pth de plusieurs centaines de Mo du dossier des logs au dossier des poids pour une inférence forcée peut entraîner des erreurs telles que des f0, tgt_sr, ou d'autres clés manquantes. Vous devez utiliser l'onglet ckpt en bas pour sélectionner manuellement ou automatiquement (si l'information se trouve dans les logs/nom_de_l'expérience), si vous souhaitez inclure les informations sur la hauteur et les options de taux d'échantillonnage audio cible, puis extraire le modèle plus petit. Après extraction, il y aura un fichier pth de 60+ MB dans le dossier des poids, et vous pouvez actualiser les voix pour l'utiliser.
+
+## Q4: Erreur de connexion.
+Il se peut que vous ayez fermé la console (fenêtre de ligne de commande noire).
+
+## Q5: WebUI affiche 'Expecting value: line 1 column 1 (char 0)'.
+Veuillez désactiver le proxy système LAN/proxy global puis rafraîchir.
+
+## Q6: Comment s'entraîner et déduire sans le WebUI ?
+Script d'entraînement :
+Vous pouvez d'abord lancer l'entraînement dans WebUI, et les versions en ligne de commande de la préparation du jeu de données et de l'entraînement seront affichées dans la fenêtre de message.
+
+Script d'inférence :
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+Par exemple :
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" récolte "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#récolte ou pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+### Explication des arguments :
+
+1. **Numéro de voix cible** : `0` (dans cet exemple)
+2. **Chemin du fichier audio d'entrée** : `"C:\ YOUR PATH FOR THE ROOT (RVC0813Nvidia)\INPUTS_VOCAL\vocal.wav"`
+3. **Chemin du fichier index** : `"C:\ YOUR PATH FOR THE ROOT (RVC0813Nvidia)\logs\Hagrid.index"`
+4. **Méthode pour l'extraction du pitch (F0)** : `harvest` (dans cet exemple)
+5. **Chemin de sortie pour le fichier audio traité** : `"C:\ YOUR PATH FOR THE ROOT (RVC0813Nvidia)\INPUTS_VOCAL\test.wav"`
+6. **Chemin du modèle** : `"C:\ YOUR PATH FOR THE ROOT (RVC0813Nvidia)\weights\HagridFR.pth"`
+7. **Taux d'index** : `0.6` (dans cet exemple)
+8. **Périphérique pour l'exécution (GPU/CPU)** : `cuda:0` pour une carte NVIDIA, par exemple.
+9. **Protection des droits d'auteur (True/False)**.
+
+
+
+## Q7: Erreur Cuda/Mémoire Cuda épuisée.
+Il y a une faible chance qu'il y ait un problème avec la configuration CUDA ou que le dispositif ne soit pas pris en charge ; plus probablement, il n'y a pas assez de mémoire (manque de mémoire).
+
+Pour l'entraînement, réduisez la taille du lot (si la réduction à 1 n'est toujours pas suffisante, vous devrez peut-être changer la carte graphique) ; pour l'inférence, ajustez les paramètres x_pad, x_query, x_center, et x_max dans le fichier config.py selon les besoins. Les cartes mémoire de 4 Go ou moins (par exemple 1060(3G) et diverses cartes de 2 Go) peuvent être abandonnées, tandis que les cartes mémoire de 4 Go ont encore une chance.
+
+## Q8: Combien de total_epoch sont optimaux ?
+Si la qualité audio du jeu d'entraînement est médiocre et que le niveau de bruit est élevé, 20-30 époques sont suffisantes. Le fixer trop haut n'améliorera pas la qualité audio de votre jeu d'entraînement de faible qualité.
+
+Si la qualité audio du jeu d'entraînement est élevée, le niveau de bruit est faible, et la durée est suffisante, vous pouvez l'augmenter. 200 est acceptable (puisque l'entraînement est rapide, et si vous êtes capable de préparer un jeu d'entraînement de haute qualité, votre GPU peut probablement gérer une durée d'entraînement plus longue sans problème).
+
+## Q9: Quelle durée de jeu d'entraînement est nécessaire ?
+Un jeu d'environ 10 min à 50 min est recommandé.
+
+Avec une garantie de haute qualité sonore et de faible bruit de fond, plus peut être ajouté si le timbre du jeu est uniforme.
+
+Pour un jeu d'entraînement de haut niveau (ton maigre + ton distinctif), 5 min à 10 min sont suffisantes.
+
+Il y a des personnes qui ont réussi à s'entraîner avec des données de 1 min à 2 min, mais le succès n'est pas reproductible par d'autres et n'est pas très informatif.
Cela nécessite que le jeu d'entraînement ait un timbre très distinctif (par exemple, un son de fille d'anime aérien à haute fréquence) et que la qualité de l'audio soit élevée ;
+Aucune tentative réussie n'a été faite jusqu'à présent avec des données de moins de 1 min. Cela n'est pas recommandé.
+
+## Q10: À quoi sert le taux d'index et comment l'ajuster ?
+Si la qualité tonale du modèle pré-entraîné et de la source d'inférence est supérieure à celle du jeu d'entraînement, ils peuvent améliorer la qualité tonale du résultat d'inférence, mais au prix d'un possible biais tonal vers le ton du modèle sous-jacent/source d'inférence plutôt que le ton du jeu d'entraînement, ce qui est généralement appelé "fuite de ton".
+
+Le taux d'index est utilisé pour réduire/résoudre le problème de la fuite de timbre. Si le taux d'index est fixé à 1, théoriquement il n'y a pas de fuite de timbre de la source d'inférence et la qualité du timbre est plus biaisée vers le jeu d'entraînement. Si le jeu d'entraînement a une qualité sonore inférieure à celle de la source d'inférence, alors un taux d'index plus élevé peut réduire la qualité sonore. Le réduire à 0 n'a pas l'effet d'utiliser le mélange de récupération pour protéger les tons du jeu d'entraînement.
+
+Si le jeu d'entraînement a une bonne qualité audio et une longue durée, augmentez le total_epoch, lorsque le modèle lui-même est moins susceptible de se référer à la source déduite et au modèle sous-jacent pré-entraîné, et qu'il y a peu de "fuite de ton", le taux d'index n'est pas important et vous pouvez même ne pas créer/partager le fichier index.
+
+## Q11: Comment choisir le gpu lors de l'inférence ?
+Dans le fichier config.py, sélectionnez le numéro de carte après "device cuda:".
+
+La correspondance entre le numéro de carte et la carte graphique peut être vue dans la section d'information de la carte graphique de l'onglet d'entraînement.
+
+## Q12: Comment utiliser le modèle sauvegardé au milieu de l'entraînement ?
+Sauvegardez via l'extraction de modèle en bas de l'onglet de traitement ckpt.
+
+## Q13: Erreur de fichier/erreur de mémoire (lors de l'entraînement) ?
+Il y a trop de processus et votre mémoire n'est pas suffisante. Vous pouvez le corriger en :
+
+1. Diminuer l'entrée dans le champ "Threads of CPU".
+
+2. Pré-découper le jeu d'entraînement en fichiers audio plus courts.
+
+## Q14: Comment poursuivre l'entraînement avec plus de données
+
+étape 1 : mettre toutes les données wav dans path2.
+
+étape 2 : exp_name2+path2 -> traiter le jeu de données et extraire la caractéristique.
+
+étape 3 : copier les derniers fichiers G et D de exp_name1 (votre expérience précédente) dans le dossier exp_name2.
+
+étape 4 : cliquez sur "entraîner le modèle", et il continuera l'entraînement depuis le début de votre époque de modèle exp précédente.
+
+## Q15: erreur à propos de llvmlite.dll
+
+OSError: Impossible de charger le fichier objet partagé : llvmlite.dll
+
+FileNotFoundError: Impossible de trouver le module lib\site-packages\llvmlite\binding\llvmlite.dll (ou l'une de ses dépendances). Essayez d'utiliser la syntaxe complète du constructeur.
+
+Le problème se produira sous Windows, installez https://aka.ms/vs/17/release/vc_redist.x64.exe et il sera corrigé.
+
+## Q16: RuntimeError: La taille étendue du tensor (17280) doit correspondre à la taille existante (0) à la dimension non-singleton 1. Tailles cibles : [1, 17280]. Tailles des tensors : [0]
+
+Supprimez les fichiers wav dont la taille est nettement inférieure à celle des autres, et cela ne se reproduira plus. Ensuite, cliquez sur "entraîner le modèle" et "entraîner l'index".
+
+## Q17: RuntimeError: La taille du tensor a (24) doit correspondre à la taille du tensor b (16) à la dimension non-singleton 2
+
+Ne changez pas le taux d'échantillonnage puis continuez l'entraînement. S'il est nécessaire de changer, le nom de l'expérience doit être modifié et le modèle sera formé à partir de zéro. Vous pouvez également copier les hauteurs et caractéristiques (dossiers 0/1/2/2b) extraites la dernière fois pour accélérer le processus d'entraînement.
+
diff --git a/docs/fr/training_tips_fr.md b/docs/fr/training_tips_fr.md
new file mode 100644
index 0000000000000000000000000000000000000000..836b45049c31741fd80efb4b1a344935c89bb878
--- /dev/null
+++ b/docs/fr/training_tips_fr.md
@@ -0,0 +1,65 @@
+Instructions et conseils pour la formation RVC
+======================================
+Ces conseils expliquent comment se déroule la formation des données.
+
+# Flux de formation
+Je vais expliquer selon les étapes de l'onglet de formation de l'interface graphique.
+
+## étape 1
+Définissez ici le nom de l'expérience.
+
+Vous pouvez également définir ici si le modèle doit prendre en compte le pitch.
+Si le modèle ne considère pas le pitch, le modèle sera plus léger, mais pas adapté au chant.
+
+Les données de chaque expérience sont placées dans `/logs/nom-de-votre-experience/`.
+
+## étape 2a
+Charge et pré-traite l'audio.
+
+### charger l'audio
+Si vous spécifiez un dossier avec de l'audio, les fichiers audio de ce dossier seront lus automatiquement.
+Par exemple, si vous spécifiez `C:Users\hoge\voices`, `C:Users\hoge\voices\voice.mp3` sera chargé, mais `C:Users\hoge\voices\dir\voice.mp3` ne sera pas chargé.
+
+Comme ffmpeg est utilisé en interne pour lire l'audio, si l'extension est prise en charge par ffmpeg, elle sera lue automatiquement.
+Après la conversion en int16 avec ffmpeg, convertir en float32 et normaliser entre -1 et 1.
+
+### débruitage
+L'audio est lissé par filtfilt de scipy.
+
+### Séparation audio
+Tout d'abord, l'audio d'entrée est divisé en détectant des parties de silence qui durent plus d'une certaine période (max_sil_kept = 5 secondes ?). Après avoir séparé l'audio sur le silence, séparez l'audio toutes les 4 secondes avec un chevauchement de 0,3 seconde. Pour l'audio séparé en 4 secondes, après normalisation du volume, convertir le fichier wav en `/logs/nom-de-votre-experience/0_gt_wavs` puis le convertir à un taux d'échantillonnage de 16k dans `/logs/nom-de-votre-experience/1_16k_wavs` sous forme de fichier wav.
+
+## étape 2b
+### Extraire le pitch
+Extrait les informations de pitch des fichiers wav. Extraire les informations de pitch (=f0) en utilisant la méthode intégrée dans parselmouth ou pyworld et les sauvegarder dans `/logs/nom-de-votre-experience/2a_f0`. Convertissez ensuite logarithmiquement les informations de pitch en un entier entre 1 et 255 et sauvegardez-le dans `/logs/nom-de-votre-experience/2b-f0nsf`.
+
+### Extraire l'empreinte de caractéristique
+Convertissez le fichier wav en incorporation à l'avance en utilisant HuBERT. Lisez le fichier wav sauvegardé dans `/logs/nom-de-votre-experience/1_16k_wavs`, convertissez le fichier wav en caractéristiques de dimension 256 avec HuBERT, et sauvegardez au format npy dans `/logs/nom-de-votre-experience/3_feature256`.
+
+## étape 3
+former le modèle.
+### Glossaire pour les débutants
+Dans l'apprentissage profond, l'ensemble de données est divisé et l'apprentissage progresse petit à petit. Dans une mise à jour de modèle (étape), les données de batch_size sont récupérées et des prédictions et corrections d'erreur sont effectuées. Faire cela une fois pour un ensemble de données compte comme une époque.
+
+Par conséquent, le temps d'apprentissage est le temps d'apprentissage par étape x (le nombre de données dans l'ensemble de données / taille du lot) x le nombre d'époques. En général, plus la taille du lot est grande, plus l'apprentissage devient stable (temps d'apprentissage par étape ÷ taille du lot) devient plus petit, mais il utilise plus de mémoire GPU. La RAM GPU peut être vérifiée avec la commande nvidia-smi. L'apprentissage peut être effectué en peu de temps en augmentant la taille du lot autant que possible selon la machine de l'environnement d'exécution.
+
+### Spécifier le modèle pré-entraîné
+RVC commence à former le modèle à partir de poids pré-entraînés plutôt que de zéro, il peut donc être formé avec un petit ensemble de données.
+
+Par défaut :
+
+- Si vous considérez le pitch, il charge `rvc-location/pretrained/f0G40k.pth` et `rvc-location/pretrained/f0D40k.pth`.
+- Si vous ne considérez pas le pitch, il charge `rvc-location/pretrained/f0G40k.pth` et `rvc-location/pretrained/f0D40k.pth`.
+
+Lors de l'apprentissage, les paramètres du modèle sont sauvegardés dans `logs/nom-de-votre-experience/G_{}.pth` et `logs/nom-de-votre-experience/D_{}.pth` pour chaque save_every_epoch, mais en spécifiant ce chemin, vous pouvez démarrer l'apprentissage. Vous pouvez redémarrer ou commencer à former à partir de poids de modèle appris lors d'une expérience différente.
+
+### Index d'apprentissage
+RVC sauvegarde les valeurs de caractéristique HuBERT utilisées lors de la formation, et pendant l'inférence, recherche les valeurs de caractéristique qui sont similaires aux valeurs de caractéristique utilisées lors de l'apprentissage pour effectuer l'inférence. Afin d'effectuer cette recherche à haute vitesse, l'index est appris à l'avance.
+Pour l'apprentissage d'index, nous utilisons la bibliothèque de recherche de voisinage approximatif faiss. Lisez la valeur de caractéristique de `logs/nom-de-votre-experience/3_feature256` et utilisez-la pour apprendre l'index, et sauvegardez-la sous `logs/nom-de-votre-experience/add_XXX.index`.
+
+(À partir de la version de mise à jour 20230428, elle est lue à partir de l'index, et la sauvegarde / spécification n'est plus nécessaire.)
+
+### Description du bouton
+- Former le modèle : après avoir exécuté l'étape 2b, appuyez sur ce bouton pour former le modèle.
+- Former l'index de caractéristique : après avoir formé le modèle, effectuez un apprentissage d'index.
+- Formation en un clic : étape 2b, formation du modèle et formation de l'index de caractéristique tout d'un coup.```
diff --git a/docs/jp/README.ja.md b/docs/jp/README.ja.md
new file mode 100644
index 0000000000000000000000000000000000000000..153a231df168cf9712230115cf9cbb5184c7981f
--- /dev/null
+++ b/docs/jp/README.ja.md
@@ -0,0 +1,203 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+VITSに基づく使いやすい音声変換（voice changer）framework
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+[**よくある質問**](./faq_ja.md) | [**AutoDLで推論(中国語のみ)**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B) | [**対照実験記録**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%AF%B9%E7%85%A7%E5%AE%9E%E9%AA%8C%C2%B7%E5%AE%9E%E9%AA%8C%E8%AE%B0%E5%BD%95) | [**オンラインデモ(中国語のみ)**](https://modelscope.cn/studios/FlowerCry/RVCv2demo)
+
+[**English**](../en/README.en.md) | [**中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+
+
+> 著作権侵害を心配することなく使用できるよう、約 50 時間の高品質なオープンソースデータセットを使用し、基底モデルを学習し出しました。
+
+> RVCv3 の基底モデルをご期待ください。より大きなパラメータ、より大きなデータ、より良い効果を提供し、基本的に同様の推論速度を維持しながら学習に必要なデータ量はより少なくなります。
+
+> モデルや統合パッケージをダウンロードしやすい[RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)のご利用がお勧めです。
+
+| 学習・推論 |
+| :--------: |
+|  |
+
+| 即時音声変換 |
+| :---------: |
+|  |
+
+## はじめに
+
+本リポジトリには下記の特徴があります。
+
+- Top1 検索を用いることで、生の特徴量を学習用データセット特徴量に変換し、トーンリーケージを削減します。
+- 比較的貧弱な GPU でも、高速かつ簡単に学習できます。
+- 少量のデータセットからでも、比較的良い結果を得ることができます。（10 分以上のノイズの少ない音声を推奨します。）
+- モデルを融合することで、音声を混ぜることができます。（ckpt processing タブの、ckpt merge を使用します。）
+- 使いやすい WebUI。
+- UVR5 Model も含んでいるため、人の声と BGM を素早く分離できます。
+- 最先端の[人間の声のピッチ抽出アルゴリズム InterSpeech2023-RMVPE](#参照プロジェクト)を使用して無声音問題を解決します。効果は最高（著しく）で、crepe_full よりも速く、リソース使用が少ないです。
+- AMD GPU と Intel GPU の加速サポート
+
+デモ動画は[こちら](https://www.bilibili.com/video/BV1pm4y1z7Gm/)でご覧ください。
+
+## 環境構築
+### Python バージョン制限
+> conda で Python 環境を管理することがお勧めです
+
+> バージョン制限の原因はこの [bug](https://github.com/facebookresearch/fairseq/issues/5012) を参照してください。
+
+```bash
+python --version # 3.8 <= Python < 3.11
+```
+
+### Linux/MacOS ワンクリック依存関係インストール・起動するスクリプト
+プロジェクトのルートディレクトリで`run.sh`を実行するだけで、`venv`仮想環境を一括設定し、必要な依存関係を自動的にインストールし、メインプログラムを起動できます。
+```bash
+sh ./run.sh
+```
+
+### 依存関係のマニュアルインストレーション
+1. `pytorch`とそのコア依存関係をインストールします。すでにインストールされている場合は見送りできます。参考: https://pytorch.org/get-started/locally/
+ ```bash
+ pip install torch torchvision torchaudio
+ ```
+2. もし、Windows + Nvidia Ampere (RTX30xx)の場合、#21 の経験に基づき、pytorchの対応する CUDA バージョンを指定する必要があります。
+ ```bash
+ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+ ```
+3. 自分の GPU に対応する依存関係をインストールします。
+- Nvidia GPU
+ ```bash
+ pip install -r requirements/main.txt
+ ```
+- AMD/Intel GPU
+ ```bash
+ pip install -r requirements/dml.txt
+ ```
+- AMD ROCM (Linux)
+ ```bash
+ pip install -r requirements/amd.txt
+ ```
+- Intel IPEX (Linux)
+ ```bash
+ pip install -r requirements/ipex.txt
+ ```
+
+## その他のデータを準備
+
+### 1. アセット
+> RVCは、`assets`フォルダにある幾つかのモデルリソースで推論・学習することが必要です。
+#### リソースの自動チェック/ダウンロード（デフォルト）
+> デフォルトでは、RVC は主プログラムの起動時に必要なリソースの完全性を自動的にチェックしできます。
+
+> リソースが不完全でも、プログラムは起動し続けます。
+
+- すべてのリソースをダウンロードしたい場合は、`--update`パラメータを追加してください。
+- 起動時のリソース完全性チェックを不要の場合は、`--nocheck`パラメータを追加してください。
+
+#### リソースのマニュアルダウンロード
+> すべてのリソースファイルは[Hugging Face space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)にあります。
+
+> `tools`フォルダでそれらをダウンロードするスクリプトを見つけることができます。
+
+> モデル/統合パッケージ/ツールの一括ダウンローダー、[RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)も使用できます。
+
+以下は、RVCが必要とするすべての事前モデルデータやその他のファイルの名前を含むリストです。
+
+- ./assets/hubert/hubert_base.pt
+ ```bash
+ rvcmd assets/hubert # RVC-Models-Downloader command
+ ```
+- ./assets/pretrained
+ ```bash
+ rvcmd assets/v1 # RVC-Models-Downloader command
+ ```
+- ./assets/uvr5_weights
+ ```bash
+ rvcmd assets/uvr5 # RVC-Models-Downloader command
+ ```
+v2バージョンのモデルを使用したい場合は、追加ダウンロードが必要です。
+
+- ./assets/pretrained_v2
+ ```bash
+ rvcmd assets/v2 # RVC-Models-Downloader command
+ ```
+
+### 2. RMVPE人声音高抽出アルゴリズムに必要なファイルのダウンロード
+
+最新のRMVPE人声音高抽出アルゴリズムを使用したい場合は、音高抽出モデルをダウンロードし、`assets/rmvpe`に配置する必要があります。
+
+- [rmvpe.pt](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+#### RMVPE(dml環境)のダウンロード（オプション、AMD/Intel GPU ユーザー）
+
+- [rmvpe.onnx](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx)
+ ```bash
+ rvcmd assets/rmvpe # RVC-Models-Downloader command
+ ```
+
+### 3. AMD ROCM（オプション、Linuxのみ）
+
+AMDのRocm技術を基にLinuxシステムでRVCを実行したい場合は、まず[ここ](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html)で必要なドライバをインストールしてください。
+
+Arch Linuxを使用している場合は、pacmanを使用して必要なドライバをインストールできます。
+````
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+````
+一部のグラフィックカードモデルでは、以下のような環境変数を追加で設定する必要があるかもしれません（例：RX6700XT）。
+````
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+````
+また、現在のユーザーが`render`および`video`ユーザーグループに所属していることを確認してください。
+````
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+````
+
+## 利用開始
+### 直接起動
+以下のコマンドで WebUI を起動します
+```bash
+python web.py
+```
+### Linux/MacOS
+```bash
+./run.sh
+```
+### IPEX 技術が必要な Intel GPU ユーザー向け(Linux のみ)
+```bash
+source /opt/intel/oneapi/setvars.sh
+./run.sh
+```
+### 統合パッケージの使用 (Windowsのみ)
+`RVC-beta.7z`をダウンロードして解凍し、`go-web.bat`をダブルクリック。
+```bash
+rvcmd packs/general/latest # RVC-Models-Downloader command
+```
+
+## 参考プロジェクト
+- [ContentVec](https://github.com/auspicious3000/contentvec/)
+- [VITS](https://github.com/jaywalnut310/vits)
+- [HIFIGAN](https://github.com/jik876/hifi-gan)
+- [Gradio](https://github.com/gradio-app/gradio)
+- [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [Vocal pitch extraction:RMVPE](https://github.com/Dream-High/RMVPE)
+ - 事前学習されたモデルは[yxlllc](https://github.com/yxlllc/RMVPE)と[RVC-Boss](https://github.com/RVC-Boss)によって学習され、テストされました。
+
+## すべての貢献者の努力に感謝します
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
diff --git a/docs/jp/faiss_tips_ja.md b/docs/jp/faiss_tips_ja.md
new file mode 100644
index 0000000000000000000000000000000000000000..89cf5ba565d6c2ccdab5c8ec9566663bfdc3fbfc
--- /dev/null
+++ b/docs/jp/faiss_tips_ja.md
@@ -0,0 +1,101 @@
+faiss tuning TIPS
+==================
+# about faiss
+faissはfacebook researchの開発する、密なベクトルに対する近傍探索をまとめたライブラリで、多くの近似近傍探索の手法を効率的に実装しています。
+近似近傍探索はある程度精度を犠牲にしながら高速に類似するベクトルを探します。
+
+## faiss in RVC
+RVCではHuBERTで変換した特徴量のEmbeddingに対し、学習データから生成されたEmbeddingと類似するものを検索し、混ぜることでより元の音声に近い変換を実現しています。ただ、この検索は愚直に行うと時間がかかるため、近似近傍探索を用いることで高速な変換を実現しています。
+
+# 実装のoverview
+モデルが配置されている '/logs/your-experiment/3_feature256'には各音声データからHuBERTで抽出された特徴量が配置されています。
+ここからnpyファイルをファイル名でソートした順番で読み込み、ベクトルを連結してbig_npyを作成しfaissを学習させます。(このベクトルのshapeは[N, 256]です。)
+
+本Tipsではまずこれらのパラメータの意味を解説します。
+
+# 手法の解説
+## index factory
+index factoryは複数の近似近傍探索の手法を繋げるパイプラインをstringで表記するfaiss独自の記法です。
+これにより、index factoryの文字列を変更するだけで様々な近似近傍探索の手法を試せます。
+RVCでは以下のように使われています。
+
+```python
+index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+index_factoryの引数のうち、1つ目はベクトルの次元数、2つ目はindex factoryの文字列で、3つ目には用いる距離を指定することができます。
+
+より詳細な記法については
+https://github.com/facebookresearch/faiss/wiki/The-index-factory
+
+## 距離指標
+embeddingの類似度として用いられる代表的な指標として以下の二つがあります。
+
+- ユークリッド距離(METRIC_L2)
+- 内積(METRIC_INNER_PRODUCT)
+
+ユークリッド距離では各次元において二乗の差をとり、全次元の差を足してから平方根をとります。これは日常的に用いる2次元、3次元での距離と同じです。
+内積はこのままでは類似度の指標として用いず、一般的にはL2ノルムで正規化してから内積をとるコサイン類似度を用います。
+
+どちらがよいかは場合によりますが、word2vec等で得られるembeddingやArcFace等で学習した類似画像検索のモデルではコサイン類似度が用いられることが多いです。ベクトルXに対してl2正規化をnumpyで行う場合は、0 divisionを避けるために十分に小さな値をepsとして以下のコードで可能です。
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+また、index factoryには第3引数に渡す値を選ぶことで計算に用いる距離指標を変更できます。
+
+```python
+index = faiss.index_factory(dimention, text, faiss.METRIC_INNER_PRODUCT)
+```
+
+## IVF
+IVF(Inverted file indexes)は全文検索における転置インデックスと似たようなアルゴリズムです。
+学習時には検索対象に対してkmeansでクラスタリングを行い、クラスタ中心を用いてボロノイ分割を行います。各データ点には一つずつクラスタが割り当てられるので、クラスタからデータ点を逆引きする辞書を作成します。
+
+例えば以下のようにクラスタが割り当てられた場合
+|index|クラスタ|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+作成される転置インデックスは以下のようになります。
+
+|クラスタ|index|
+|-------|-----|
+|A|1, 3|
+|B|2, 5|
+|C|4|
+
+検索時にはまずクラスタからn_probe個のクラスタを検索し、次にそれぞれのクラスタに属するデータ点について距離を計算します。
+
+# 推奨されるパラメータ
+indexの選び方については公式にガイドラインがあるので、それに準じて説明します。
+https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
+
+1M以下のデータセットにおいては4bit-PQが2023年4月時点ではfaissで利用できる最も効率的な手法です。
+これをIVFと組み合わせ、4bit-PQで候補を絞り、最後に正確な指標で距離を再計算するには以下のindex factoryを用いることで記載できます。
+
+```python
+index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## IVFの推奨パラメータ
+IVFの数が多すぎる場合、たとえばデータ数の数だけIVFによる粗量子化を行うと、これは愚直な全探索と同じになり効率が悪いです。
+1M以下の場合ではIVFの値はデータ点の数Nに対して4*sqrt(N) ~ 16*sqrt(N)に推奨しています。
+
+n_probeはn_probeの数に比例して計算時間が増えるので、精度と相談して適切に選んでください。個人的にはRVCにおいてそこまで精度は必要ないと思うのでn_probe = 1で良いと思います。
+
+## FastScan
+FastScanは直積量子化で大まかに距離を近似するのを、レジスタ内で行うことにより高速に行うようにした手法です。
+直積量子化は学習時にd次元ごと(通常はd=2)に独立してクラスタリングを行い、クラスタ同士の距離を事前計算してlookup tableを作成します。予測時はlookup tableを見ることで各次元の距離をO(1)で計算できます。
+そのため、PQの次に指定する数字は通常ベクトルの半分の次元を指定します。
+
+FastScanに関するより詳細な説明は公式のドキュメントを参照してください。
+https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlatはFastScanで計算した大まかな距離を、index factoryの第三引数で指定した正確な距離で再計算する指示です。
+k個の近傍を取得する際は、k*k_factor個の点について再計算が行われます。
diff --git a/docs/jp/faq_ja.md b/docs/jp/faq_ja.md
new file mode 100644
index 0000000000000000000000000000000000000000..db6935defd5fe7c5a3c752d6dc9193fe67e756b5
--- /dev/null
+++ b/docs/jp/faq_ja.md
@@ -0,0 +1,117 @@
+## Q1: ワンクリックトレーニングが終わってもインデックスがない
+
+"Training is done. The program is closed."と表示された場合、モデルトレーニングは成功しています。その直後のエラーは誤りです。
+
+ワンクリックトレーニングが終了しても added で始まるインデックスファイルがない場合、トレーニングセットが大きすぎてインデックス追加のステップが停止している可能性があります。バッチ処理 add インデックスでメモリの要求が高すぎる問題を解決しました。一時的に「トレーニングインデックス」ボタンをもう一度クリックしてみてください。
+
+## Q2: トレーニングが終了してもトレーニングセットの音色が見えない
+
+音色をリフレッシュしてもう一度確認してください。それでも見えない場合は、トレーニングにエラーがなかったか、コンソールと WebUI のスクリーンショット、logs/実験名の下のログを開発者に送って確認してみてください。
+
+## Q3: モデルをどのように共有するか
+
+rvc_root/logs/実験名の下に保存されている pth は、推論に使用するために共有するためのものではなく、実験の状態を保存して再現およびトレーニングを続けるためのものです。共有するためのモデルは、weights フォルダの下にある 60MB 以上の pth ファイルです。
+ 今後、weights/exp_name.pth と logs/exp_name/added_xxx.index を組み合わせて weights/exp_name.zip にパッケージ化し、インデックスの記入ステップを省略します。その場合、zip ファイルを共有し、pth ファイルは共有しないでください。別のマシンでトレーニングを続ける場合を除きます。
+ logs フォルダの数百 MB の pth ファイルを weights フォルダにコピー/共有して推論に強制的に使用すると、f0、tgt_sr などのさまざまなキーが存在しないというエラーが発生する可能性があります。ckpt タブの一番下で、音高、目標オーディオサンプリングレートを手動または自動（ローカルの logs に関連情報が見つかる場合は自動的に）で選択してから、ckpt の小型モデルを抽出する必要があります（入力パスに G で始まるものを記入）。抽出が完了すると、weights フォルダに 60MB 以上の pth ファイルが表示され、音色をリフレッシュした後に使用できます。
+
+## Q4: Connection Error
+
+コンソール（黒いウィンドウ）を閉じた可能性があります。
+
+## Q5: WebUI が Expecting value: line 1 column 1 (char 0)と表示する
+
+システムのローカルネットワークプロキシ/グローバルプロキシを閉じてください。
+
+これはクライアントのプロキシだけでなく、サーバー側のプロキシも含まれます（例えば autodl で http_proxy と https_proxy を設定して学術的な加速を行っている場合、使用する際には unset でオフにする必要があります）。
+
+## Q6: WebUI を使わずにコマンドでトレーニングや推論を行うには
+
+トレーニングスクリプト：
+まず WebUI を実行し、メッセージウィンドウにデータセット処理とトレーニング用のコマンドラインが表示されます。
+
+推論スクリプト：
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+例：
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#harvest or pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+## Q7: Cuda error/Cuda out of memory
+
+まれに cuda の設定問題やデバイスがサポートされていない可能性がありますが、大半はメモリ不足（out of memory）が原因です。
+
+トレーニングの場合は batch size を小さくします（1 にしても足りない場合はグラフィックカードを変更するしかありません）。推論の場合は、config.py の末尾にある x_pad、x_query、x_center、x_max を適宜小さくします。4GB 以下のメモリ（例えば 1060（3G）や各種 2GB のグラフィックカード）は諦めることをお勧めしますが、4GB のメモリのグラフィックカードはまだ救いがあります。
+
+## Q8: total_epoch はどのくらいに設定するのが良いですか
+
+トレーニングセットの音質が悪く、ノイズが多い場合は、20〜30 で十分です。高すぎると、ベースモデルの音質が低音質のトレーニングセットを高めることができません。
+トレーニングセットの音質が高く、ノイズが少なく、長い場合は、高く設定できます。200 は問題ありません（トレーニング速度が速いので、高音質のトレーニングセットを準備できる条件がある場合、グラフィックカードも条件が良いはずなので、少しトレーニング時間が長くなることを気にすることはありません）。
+
+## Q9: トレーニングセットはどれくらいの長さが必要ですか
+
+10 分から 50 分を推奨します。
+ 音質が良く、バックグラウンドノイズが低い場合、個人的な特徴のある音色であれば、多ければ多いほど良いです。
+ 高品質のトレーニングセット（精巧に準備された + 特徴的な音色）であれば、5 分から 10 分でも大丈夫です。リポジトリの作者もよくこの方法で遊びます。
+ 1 分から 2 分のデータでトレーニングに成功した人もいますが、その成功体験は他人には再現できないため、あまり参考になりません。トレーニングセットの音色が非常に特徴的である必要があります（例：高い周波数の透明な声や少女の声など）、そして音質が良い必要があります。
+ 1 分未満のデータでトレーニングを試みた（成功した）ケースはまだ見たことがありません。このような試みはお勧めしません。
+
+## Q10: index rate は何に使うもので、どのように調整するのか（啓蒙）
+
+もしベースモデルや推論ソースの音質がトレーニングセットよりも高い場合、推論結果の音質を向上させることができますが、音色がベースモデル/推論ソースの音色に近づくことがあります。これを「音色漏れ」と言います。
+ index rate は音色漏れの問題を減少させたり解決するために使用されます。1 に設定すると、理論的には推論ソースの音色漏れの問題は存在しませんが、音質はトレーニングセットに近づきます。トレーニングセットの音質が推論ソースよりも低い場合、index rate を高くすると音質が低下する可能性があります。0 に設定すると、検索ミックスを利用してトレーニングセットの音色を保護する効果はありません。
+ トレーニングセットが高品質で長い場合、total_epoch を高く設定することができ、この場合、モデル自体は推論ソースやベースモデルの音色をあまり参照しないため、「音色漏れ」の問題はほとんど発生しません。この時、index rate は重要ではなく、インデックスファイルを作成したり共有したりする必要もありません。
+
+## Q11: 推論時に GPU をどのように選択するか
+
+config.py ファイルの device cuda:の後にカード番号を選択します。
+カード番号とグラフィックカードのマッピング関係は、トレーニングタブのグラフィックカード情報欄で確認できます。
+
+## Q12: トレーニング中に保存された pth ファイルをどのように推論するか
+
+ckpt タブの一番下で小型モデルを抽出します。
+
+## Q13: トレーニングをどのように中断し、続行するか
+
+現在の段階では、WebUI コンソールを閉じて go-web.bat をダブルクリックしてプログラムを再起動するしかありません。ウェブページのパラメータもリフレッシュして再度入力する必要があります。
+トレーニングを続けるには：同じウェブページのパラメータでトレーニングモデルをクリックすると、前回のチェックポイントからトレーニングを続けます。
+
+## Q14: トレーニング中にファイルページ/メモリエラーが発生した場合の対処法
+
+プロセスが多すぎてメモリがオーバーフローしました。以下の方法で解決できるかもしれません。
+
+1. 「音高抽出とデータ処理に使用する CPU プロセス数」を適宜下げます。
+2. トレーニングセットのオーディオを手動でカットして、あまり長くならないようにします。
+
+## Q15: 途中でデータを追加してトレーニングする方法
+
+1. 全データに新しい実験名を作成します。
+2. 前回の最新の G と D ファイル（あるいはどの中間 ckpt を基にトレーニングしたい場合は、その中間のものをコピーすることもできます）を新しい実験名にコピーします。
+3. 新しい実験名でワンクリックトレーニングを開始すると、前回の最新の進捗からトレーニングを続けます。
+
+## Q16: llvmlite.dll に関するエラー
+
+```bash
+OSError: Could not load shared object file: llvmlite.dll
+
+FileNotFoundError: Could not find module lib\site-packages\llvmlite\binding\llvmlite.dll (or one of its dependencies). Try using the full path with constructor syntax.
+```
+
+Windows プラットフォームではこのエラーが発生しますが、https://aka.ms/vs/17/release/vc_redist.x64.exeをインストールしてWebUIを再起動すれば解決します。
+
+## Q17: RuntimeError: テンソルの拡張サイズ（17280）は、非シングルトン次元 1 での既存サイズ（0）と一致する必要があります。 ターゲットサイズ：[1, 17280]。 テンソルサイズ：[0]
+
+wavs16k フォルダーの下で、他のファイルよりも明らかに小さいいくつかのオーディオファイルを見つけて削除し、トレーニングモデルをクリックすればエラーは発生しませんが、ワンクリックプロセスが中断されたため、モデルのトレーニングが完了したらインデックスのトレーニングをクリックする必要があります。
+
+## Q18: RuntimeError: テンソル a のサイズ（24）は、非シングルトン次元 2 でテンソル b（16）のサイズと一致する必要があります
+
+トレーニング中にサンプリングレートを変更してはいけません。変更する必要がある場合は、実験名を変更して最初からトレーニングする必要があります。もちろん、前回抽出した音高と特徴（0/1/2/2b フォルダ）をコピーしてトレーニングプロセスを加速することもできます。
diff --git a/docs/jp/training_tips_ja.md b/docs/jp/training_tips_ja.md
new file mode 100644
index 0000000000000000000000000000000000000000..c5b06f2fdaa603a690c51ee2b79daecc4305fbd5
--- /dev/null
+++ b/docs/jp/training_tips_ja.md
@@ -0,0 +1,64 @@
+RVCの訓練における説明、およびTIPS
+===============================
+本TIPSではどのようにデータの訓練が行われているかを説明します。
+
+# 訓練の流れ
+GUIの訓練タブのstepに沿って説明します。
+
+## step1
+実験名の設定を行います。
+
+また、モデルに音高ガイド(ピッチ)を考慮させるかもここで設定できます。考慮させない場合はモデルは軽量になりますが、歌唱には向かなくなります。
+
+各実験のデータは`/logs/実験名/`に配置されます。
+
+## step2a
+音声の読み込みと前処理を行います。
+
+### load audio
+音声のあるフォルダを指定すると、そのフォルダ内にある音声ファイルを自動で読み込みます。
+例えば`C:Users\hoge\voices`を指定した場合、`C:Users\hoge\voices\voice.mp3`は読み込まれますが、`C:Users\hoge\voices\dir\voice.mp3`は読み込まれません。
+
+音声の読み込みには内部でffmpegを利用しているので、ffmpegで対応している拡張子であれば自動的に読み込まれます。
+ffmpegでint16に変換した後、float32に変換し、-1 ~ 1の間に正規化されます。
+
+### denoising
+音声についてscipyのfiltfiltによる平滑化を行います。
+
+### 音声の分割
+入力した音声はまず、一定期間(max_sil_kept=5秒?)より長く無音が続く部分を検知して音声を分割します。無音で音声を分割した後は、0.3秒のoverlapを含む4秒ごとに音声を分割します。4秒以内に区切られた音声は、音量の正規化を行った後wavファイルを`/logs/実験名/0_gt_wavs`に、そこから16kのサンプリングレートに変換して`/logs/実験名/1_16k_wavs`にwavファイルで保存します。
+
+## step2b
+### ピッチの抽出
+wavファイルからピッチ(音の高低)の情報を抽出します。parselmouthやpyworldに内蔵されている手法でピッチ情報(=f0)を抽出し、`/logs/実験名/2a_f0`に保存します。その後、ピッチ情報を対数で変換して1~255の整数に変換し、`/logs/実験名/2b-f0nsf`に保存します。
+
+### feature_printの抽出
+HuBERTを用いてwavファイルを事前にembeddingに変換します。`/logs/実験名/1_16k_wavs`に保存したwavファイルを読み込み、HuBERTでwavファイルを256次元の特徴量に変換し、npy形式で`/logs/実験名/3_feature256`に保存します。
+
+## step3
+モデルのトレーニングを行います。
+### 初心者向け用語解説
+深層学習ではデータセットを分割し、少しずつ学習を進めていきます。一回のモデルの更新(step)では、batch_size個のデータを取り出し予測と誤差の修正を行います。これをデータセットに対して一通り行うと一epochと数えます。
+
+そのため、学習時間は 1step当たりの学習時間 x (データセット内のデータ数 ÷ バッチサイズ) x epoch数 かかります。一般にバッチサイズを大きくするほど学習は安定し、(1step当たりの学習時間÷バッチサイズ)は小さくなりますが、その分GPUのメモリを多く使用します。GPUのRAMはnvidia-smiコマンド等で確認できます。実行環境のマシンに合わせてバッチサイズをできるだけ大きくするとより短時間で学習が可能です。
+
+### pretrained modelの指定
+RVCではモデルの訓練を0からではなく、事前学習済みの重みから開始するため、少ないデータセットで学習を行えます。
+
+デフォルトでは
+
+- 音高ガイドを考慮する場合、`RVCのある場所/pretrained/f0G40k.pth`と`RVCのある場所/pretrained/f0D40k.pth`を読み込みます。
+- 音高ガイドを考慮しない場合、`RVCのある場所/pretrained/G40k.pth`と`RVCのある場所/pretrained/D40k.pth`を読み込みます。
+
+学習時はsave_every_epochごとにモデルのパラメータが`logs/実験名/G_{}.pth`と`logs/実験名/D_{}.pth`に保存されますが、このパスを指定することで学習を再開したり、もしくは違う実験で学習したモデルの重みから学習を開始できます。
+
+### indexの学習
+RVCでは学習時に使われたHuBERTの特徴量を保存し、推論時は学習時の特徴量から近い特徴量を探してきて推論を行います。この検索を高速に行うために事前にindexの学習を行います。
+indexの学習には近似近傍探索ライブラリのfaissを用います。`/logs/実験名/3_feature256`の特徴量を読み込み、それを用いて学習したindexを`/logs/実験名/add_XXX.index`として保存します。
+(20230428updateよりtotal_fea.npyはindexから読み込むので不要になりました。)
+
+### ボタンの説明
+- モデルのトレーニング: step2bまでを実行した後、このボタンを押すとモデルの学習を行います。
+- 特徴インデックスのトレーニング: モデルのトレーニング後、indexの学習を行います。
+- ワンクリックトレーニング: step2bまでとモデルのトレーニング、特徴インデックスのトレーニングを一括で行います。
+
diff --git a/docs/kr/README.ko.han.md b/docs/kr/README.ko.han.md
new file mode 100644
index 0000000000000000000000000000000000000000..3adb565dd41e54e176fa6ad7e96640844a3de52a
--- /dev/null
+++ b/docs/kr/README.ko.han.md
@@ -0,0 +1,100 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+VITS基盤의 簡單하고使用하기 쉬운音聲變換틀
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+
+
+------
+
+[**English**](../en/README.en.md) | [**中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+> [示範映像](https://www.bilibili.com/video/BV1pm4y1z7Gm/)을 確認해 보세요!
+
+> RVC를活用한實時間音聲變換: [w-okada/voice-changer](https://github.com/w-okada/voice-changer)
+
+> 基本모델은 50時間假量의 高品質 오픈 소스 VCTK 데이터셋을 使用하였으므로, 著作權上의 念慮가 없으니 安心하고 使用하시기 바랍니다.
+
+> 著作權問題가 없는 高品質의 노래를 以後에도 繼續해서 訓練할 豫定입니다.
+
+## 紹介
+本Repo는 다음과 같은 特徵을 가지고 있습니다:
++ top1檢索을利用하여 入力音色特徵을 訓練세트音色特徵으로 代替하여 音色의漏出을 防止;
++ 相對的으로 낮은性能의 GPU에서도 빠른訓練可能;
++ 적은量의 데이터로 訓練해도 좋은 結果를 얻을 수 있음 (最小10分以上의 低雜음音聲데이터를 使用하는 것을 勸獎);
++ 모델融合을通한 音色의 變調可能 (ckpt處理탭->ckpt混合選擇);
++ 使用하기 쉬운 WebUI (웹 使用者인터페이스);
++ UVR5 모델을 利用하여 목소리와 背景音樂의 빠른 分離;
+
+## 環境의準備
+poetry를通해 依存를設置하는 것을 勸獎합니다.
+
+다음命令은 Python 버전3.8以上의環境에서 實行되어야 합니다:
+```bash
+# PyTorch 關聯主要依存設置, 이미設置되어 있는 境遇 건너뛰기 可能
+# 參照: https://pytorch.org/get-started/locally/
+pip install torch torchvision torchaudio
+
+# Windows + Nvidia Ampere Architecture(RTX30xx)를 使用하고 있다面, #21 에서 명시된 것과 같이 PyTorch에 맞는 CUDA 버전을 指定해야 합니다.
+#pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+
+# Poetry 設置, 이미設置되어 있는 境遇 건너뛰기 可能
+# Reference: https://python-poetry.org/docs/#installation
+curl -sSL https://install.python-poetry.org | python3 -
+
+# 依存設置
+poetry install
+```
+pip를 活用하여依存를 設置하여도 無妨합니다.
+
+```bash
+pip install -r requirements/main.txt
+```
+
+## 其他預備모델準備
+RVC 모델은 推論과訓練을 依하여 다른 預備모델이 必要합니다.
+
+[Huggingface space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)를 通해서 다운로드 할 수 있습니다.
+
+다음은 RVC에 必要한 預備모델 및 其他 파일 目錄입니다:
+```bash
+./assets/hubert/hubert_base.pt
+
+./assets/pretrained
+
+./assets/uvr5_weights
+
+V2 버전 모델을 테스트하려면 추가 다운로드가 필요합니다.
+
+./assets/pretrained_v2
+
+```
+그後 以下의 命令을 使用하여 WebUI를 始作할 수 있습니다:
+```bash
+python web.py
+```
+Windows를 使用하는境遇 `RVC-beta.7z`를 다운로드 및 壓縮解除하여 RVC를 直接使用하거나 `go-web.bat`을 使用하여 WebUi를 直接할 수 있습니다.
+
+## 參考
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
+## 모든寄與者분들의勞力에感謝드립니다
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
+
diff --git a/docs/kr/README.ko.md b/docs/kr/README.ko.md
new file mode 100644
index 0000000000000000000000000000000000000000..00bc807218a50190aa773cfefa910fed415fb698
--- /dev/null
+++ b/docs/kr/README.ko.md
@@ -0,0 +1,359 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+VITS 기반의 간단하고 사용하기 쉬운 음성 변환 프레임워크.
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+[**자주 묻는 질문**](./faq_ko.md) | [**AutoDL·5원으로 AI 가수 훈련**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/Autodl%E8%AE%AD%E7%BB%83RVC%C2%B7AI%E6%AD%8C%E6%89%8B%E6%95%99%E7%A8%8B) | [**대조 실험 기록**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/%E5%AF%B9%E7%85%A7%E5%AE%9E%E9%AA%8C%C2%B7%E5%AE%9E%E9%AA%8C%E8%AE%B0%E5%BD%95) | [**온라인 데모**](https://modelscope.cn/studios/FlowerCry/RVCv2demo)
+
+[**English**](../en/README.en.md) | [**中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+
+
+> 기본 모델은 50시간 가량의 고퀄리티 오픈 소스 VCTK 데이터셋을 사용하였으므로, 저작권상의 염려가 없으니 안심하고 사용하시기 바랍니다.
+
+> 더 큰 매개변수, 더 큰 데이터, 더 나은 효과, 기본적으로 동일한 추론 속도, 더 적은 양의 훈련 데이터가 필요한 RVCv3의 기본 모델을 기대해 주십시오.
+
+> 특정 지역에서 Hugging Face에 직접 연결할 수 없는 경우가 있으며, 성공적으로 연결해도 속도가 매우 느릴 수 있으므로, 모델/통합 패키지/도구의 일괄 다운로더를 특별히 소개합니다. [RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)
+
+| 훈련 및 추론 인터페이스 |
+| :--------: |
+|  |
+
+| 실시간 음성 변환 인터페이스 |
+| :---------: |
+|  |
+
+## 소개
+
+본 프로젝트는 다음과 같은 특징을 가지고 있습니다:
+
+- top1 검색을 이용하여 입력 음색 특징을 훈련 세트 음색 특징으로 대체하여 음색의 누출을 방지
+- 상대적으로 낮은 성능의 GPU에서도 빠른 훈련 가능
+- 적은 양의 데이터로 훈련해도 좋은 결과를 얻을 수 있음 (최소 10분 이상의 저잡음 음성 데이터를 사용하는 것을 권장)
+- 모델 융합을 통한 음색의 변조 가능 (ckpt 처리 탭->ckpt 병합 선택)
+- 사용하기 쉬운 WebUI (웹 인터페이스)
+- UVR5 모델을 이용하여 목소리와 배경음악의 빠른 분리;
+- 최첨단 [음성 피치 추출 알고리즘 InterSpeech2023-RMVPE](#参考项目)을 사용하여 무성음 문제를 해결합니다. 효과는 최고(압도적)이며 crepe_full보다 더 빠르고 리소스 사용이 적음
+- A카드와 I카드 가속을 지원
+
+해당 프로젝트의 [데모 비디오](https://www.bilibili.com/video/BV1pm4y1z7Gm/)를 확인해보세요!
+
+## 환경 설정
+
+다음 명령은 Python 버전이 3.8 이상인 환경에서 실행해야 합니다.
+
+### Windows/Linux/MacOS 등 플랫폼 공통 방법
+
+아래 방법 중 하나를 선택하세요.
+
+#### 1. pip를 통한 의존성 설치
+
+1. Pytorch 및 의존성 모듈 설치, 이미 설치되어 있으면 생략. 참조: https://pytorch.org/get-started/locally/
+
+```bash
+pip install torch torchvision torchaudio
+```
+
+2. win 시스템 + Nvidia Ampere 아키텍처(RTX30xx) 사용 시, #21의 사례에 따라 pytorch에 해당하는 cuda 버전을 지정
+
+```bash
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+```
+
+3. 자신의 그래픽 카드에 맞는 의존성 설치
+
+- N카드
+
+```bash
+pip install -r requirements/main.txt
+```
+
+- A카드/I카드
+
+```bash
+pip install -r requirements/dml.txt
+```
+
+- A카드ROCM(Linux)
+
+```bash
+pip install -r requirements/amd.txt
+```
+
+- I카드IPEX(Linux)
+
+```bash
+pip install -r requirements/ipex.txt
+```
+
+#### 2. poetry를 통한 의존성 설치
+
+Poetry 의존성 관리 도구 설치, 이미 설치된 경우 생략. 참조: https://python-poetry.org/docs/#installation
+
+```bash
+curl -sSL https://install.python-poetry.org | python3 -
+```
+
+poetry를 통한 의존성 설치
+
+```bash
+poetry install
+```
+
+### MacOS
+
+`run.sh`를 통해 의존성 설치 가능
+
+```bash
+sh ./run.sh
+```
+
+
+
+## 기타 사전 훈련된 모델 준비
+
+### assets
+
+> RVC는 추론과 훈련을 위해 assets 폴더 하위에 사전 훈련된 모델이 필요합니다.
+
+#### 자동 검사/다운로드 리소스(기본값)
+
+> 기본적으로 RVC는 시작할 때 필요한 리소스의 무결성을 자동으로 확인할 수 있습니다.
+
+> 리소스가 불완전하더라도 프로그램은 계속 실행됩니다.
+
+- 모든 리소스를 다운로드하려면 `--update` 매개변수를 추가하세요
+- 시작 시 리소스 무결성 검사를 건너뛰려면 `--nocheck` 매개변수를 추가하세요
+
+#### 리소스 수동 다운로드
+
+> 모든 리소스 파일은 [Hugging Face space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)에 있습니다.
+
+> 이들을 다운로드하는 스크립트는 `tools` 폴더에서 찾을 수 있습니다.
+
+> 모델/통합 패키지/도구의 일괄 다운로더를 사용할 수도 있습니다: [RVC-Models-Downloader](https://github.com/fumiama/RVC-Models-Downloader)
+
+다음은 RVC에 필요한 모든 사전 훈련된 모델과 기타 파일의 목록입니다.
+
+- ./assets/hubert/hubert_base.pt
+
+ ```bash
+ rvcmd assets/hubert # RVC-Models-Downloader command
+ ```
+
+- ./assets/pretrained
+
+ ```bash
+ rvcmd assets/v1 # RVC-Models-Downloader command
+ ```
+
+- ./assets/uvr5_weights
+ ```bash
+ rvcmd assets/uvr5 # RVC-Models-Downloader command
+ ```
+
+v2 버전 모델을 사용하려면 추가로 다음을 다운로드해야 합니다.
+
+- ./assets/pretrained_v2
+ ```bash
+ rvcmd assets/v2 # RVC-Models-Downloader command
+ ```
+
+### 2. RMVPE 인간 음성 피치 추출 알고리즘에 필요한 파일 다운로드
+
+최신 RMVPE 인간 음성 피치 추출 알고리즘을 사용하려면 음피치 추출 모델 매개변수를 다운로드하고 RVC 루트 디렉토리에 배치해야 합니다.
+
+- [rmvpe.pt 다운로드](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt)
+
+#### dml 환경의 RMVPE 다운로드(선택사항, A카드/I카드 사용자)
+
+- [rmvpe.onnx 다운로드](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx)
+
+### 3. AMD 그래픽 카드 Rocm(선택사항, Linux만 해당)
+
+Linux 시스템에서 AMD의 Rocm 기술을 기반으로 RVC를 실행하려면 [여기](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html)에서 필요한 드라이버를 먼저 설치하세요.
+
+Arch Linux를 사용하는 경우 pacman을 사용하여 필요한 드라이버를 설치할 수 있습니다.
+
+```
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+```
+
+일부 모델의 그래픽 카드(예: RX6700XT)의 경우, 다음과 같은 환경 변수를 추가로 설정해야 할 수 있습니다.
+
+```
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+```
+
+동시에 현재 사용자가 `render` 및 `video` 사용자 그룹에 속해 있는지 확인하세요.
+
+```
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+```
+
+## 시작하기
+
+### 직접 시작
+
+다음 명령어로 WebUI를 시작하세요
+
+```bash
+python web.py
+```
+
+### 통합 패키지 사용
+
+`RVC-beta.7z`를 다운로드하고 압축 해제
+
+#### Windows 사용자
+
+`go-web.bat` 더블 클릭
+
+#### MacOS 사용자
+
+```bash
+sh ./run.sh
+```
+
+### IPEX 기술이 필요한 I카드 사용자를 위한 지침(Linux만 해당)
+
+```bash
+source /opt/intel/oneapi/setvars.sh
+```
+
+## 참조 프로젝트
+
+- [ContentVec](https://github.com/auspicious3000/contentvec/)
+- [VITS](https://github.com/jaywalnut310/vits)
+- [HIFIGAN](https://github.com/jik876/hifi-gan)
+- [Gradio](https://github.com/gradio-app/gradio)
+- [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [Vocal pitch extraction:RMVPE](https://github.com/Dream-High/RMVPE)
+ - 사전 훈련된 모델은 [yxlllc](https://github.com/yxlllc/RMVPE)와 [RVC-Boss](https://github.com/RVC-Boss)에 의해 훈련되고 테스트되었습니다.
+
+## 모든 기여자들의 노력에 감사드립니다
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
diff --git a/docs/kr/faiss_tips_ko.md b/docs/kr/faiss_tips_ko.md
new file mode 100644
index 0000000000000000000000000000000000000000..9e17fb3bae9210b022fe5ce45f5d1f900f770ca7
--- /dev/null
+++ b/docs/kr/faiss_tips_ko.md
@@ -0,0 +1,132 @@
+Facebook AI Similarity Search (Faiss) 팁
+==================
+# Faiss에 대하여
+Faiss 는 Facebook Research가 개발하는, 고밀도 벡터 이웃 검색 라이브러리입니다. 근사 근접 탐색법 (Approximate Neigbor Search)은 약간의 정확성을 희생하여 유사 벡터를 고속으로 찾습니다.
+
+## RVC에 있어서 Faiss
+RVC에서는 HuBERT로 변환한 feature의 embedding을 위해 훈련 데이터에서 생성된 embedding과 유사한 embadding을 검색하고 혼합하여 원래의 음성에 더욱 가까운 변환을 달성합니다. 그러나, 이 탐색법은 단순히 수행하면 시간이 다소 소모되므로, 근사 근접 탐색법을 통해 고속 변환을 가능케 하고 있습니다.
+
+# 구현 개요
+모델이 위치한 `/logs/your-experiment/3_feature256`에는 각 음성 데이터에서 HuBERT가 추출한 feature들이 있습니다. 여기에서 파일 이름별로 정렬된 npy 파일을 읽고, 벡터를 연결하여 big_npy ([N, 256] 모양의 벡터) 를 만듭니다. big_npy를 `/logs/your-experiment/total_fea.npy`로 저장한 후, Faiss로 학습시킵니다.
+
+2023/04/18 기준으로, Faiss의 Index Factory 기능을 이용해, L2 거리에 근거하는 IVF를 이용하고 있습니다. IVF의 분할수(n_ivf)는 N//39로, n_probe는 int(np.power(n_ivf, 0.3))가 사용되고 있습니다. (web.py의 train_index 주위를 찾으십시오.)
+
+이 팁에서는 먼저 이러한 매개 변수의 의미를 설명하고, 개발자가 추후 더 나은 index를 작성할 수 있도록 하는 조언을 작성합니다.
+
+# 방법의 설명
+## Index factory
+index factory는 여러 근사 근접 탐색법을 문자열로 연결하는 pipeline을 문자열로 표기하는 Faiss만의 독자적인 기법입니다. 이를 통해 index factory의 문자열을 변경하는 것만으로 다양한 근사 근접 탐색을 시도해 볼 수 있습니다. RVC에서는 다음과 같이 사용됩니다:
+
+```python
+index = Faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+`index_factory`의 인수들 중 첫 번째는 벡터의 차원 수이고, 두번째는 index factory 문자열이며, 세번째에는 사용할 거리를 지정할 수 있습니다.
+
+기법의 보다 자세한 설명은 https://github.com/facebookresearch/Faiss/wiki/The-index-factory 를 확인해 주십시오.
+
+## 거리에 대한 index
+embedding의 유사도로서 사용되는 대표적인 지표로서 이하의 2개가 있습니다.
+
+- 유클리드 거리 (METRIC_L2)
+- 내적(内積) (METRIC_INNER_PRODUCT)
+
+유클리드 거리에서는 각 차원에서 제곱의 차를 구하고, 각 차원에서 구한 차를 모두 더한 후 제곱근을 취합니다. 이것은 일상적으로 사용되는 2차원, 3차원에서의 거리의 연산법과 같습니다. 내적은 그 값을 그대로 유사도 지표로 사용하지 않고, L2 정규화를 한 이후 내적을 취하는 코사인 유사도를 사용합니다.
+
+어느 쪽이 더 좋은지는 경우에 따라 다르지만, word2vec에서 얻은 embedding 및 ArcFace를 활용한 이미지 검색 모델은 코사인 유사성이 이용되는 경우가 많습니다. numpy를 사용하여 벡터 X에 대해 L2 정규화를 하고자 하는 경우, 0 division을 피하기 위해 충분히 작은 값을 eps로 한 뒤 이하에 코드를 활용하면 됩니다.
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+또한, `index factory`의 3번째 인수에 건네주는 값을 선택하는 것을 통해 계산에 사용하는 거리 index를 변경할 수 있습니다.
+
+```python
+index = Faiss.index_factory(dimention, text, Faiss.METRIC_INNER_PRODUCT)
+```
+
+## IVF
+IVF (Inverted file indexes)는 역색인 탐색법과 유사한 알고리즘입니다. 학습시에는 검색 대상에 대해 k-평균 군집법을 실시하고 클러스터 중심을 이용해 보로노이 분할을 실시합니다. 각 데이터 포인트에는 클러스터가 할당되므로, 클러스터에서 데이터 포인트를 조회하는 dictionary를 만듭니다.
+
+예를 들어, 클러스터가 다음과 같이 할당된 경우
+|index|Cluster|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+IVF 이후의 결과는 다음과 같습니다:
+
+|cluster|index|
+|-------|-----|
+|A|1, 3|
+|B|2, 5|
+|C|4|
+
+탐색 시, 우선 클러스터에서 `n_probe`개의 클러스터를 탐색한 다음, 각 클러스터에 속한 데이터 포인트의 거리를 계산합니다.
+
+# 권장 매개변수
+index의 선택 방법에 대해서는 공식적으로 가이드 라인이 있으므로, 거기에 준해 설명합니다.
+https://github.com/facebookresearch/Faiss/wiki/Guidelines-to-choose-an-index
+
+1M 이하의 데이터 세트에 있어서는 4bit-PQ가 2023년 4월 시점에서는 Faiss로 이용할 수 있는 가장 효율적인 수법입니다. 이것을 IVF와 조합해, 4bit-PQ로 후보를 추려내고, 마지막으로 이하의 index factory를 이용하여 정확한 지표로 거리를 재계산하면 됩니다.
+
+```python
+index = Faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## IVF 권장 매개변수
+IVF의 수가 너무 많으면, 가령 데이터 수의 수만큼 IVF로 양자화(Quantization)를 수행하면, 이것은 완전탐색과 같아져 효율이 나빠지게 됩니다. 1M 이하의 경우 IVF 값은 데이터 포인트 수 N에 대해 4sqrt(N) ~ 16sqrt(N)를 사용하는 것을 권장합니다.
+
+n_probe는 n_probe의 수에 비례하여 계산 시간이 늘어나므로 정확도와 시간을 적절히 균형을 맞추어 주십시오. 개인적으로 RVC에 있어서 그렇게까지 정확도는 필요 없다고 생각하기 때문에 n_probe = 1이면 된다고 생각합니다.
+
+## FastScan
+FastScan은 직적 양자화를 레지스터에서 수행함으로써 거리의 고속 근사를 가능하게 하는 방법입니다.직적 양자화는 학습시에 d차원마다(보통 d=2)에 독립적으로 클러스터링을 실시해, 클러스터끼리의 거리를 사전 계산해 lookup table를 작성합니다. 예측시는 lookup table을 보면 각 차원의 거리를 O(1)로 계산할 수 있습니다. 따라서 PQ 다음에 지정하는 숫자는 일반적으로 벡터의 절반 차원을 지정합니다.
+
+FastScan에 대한 자세한 설명은 공식 문서를 참조하십시오.
+https://github.com/facebookresearch/Faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlat은 FastScan이 계산한 대략적인 거리를 index factory의 3번째 인수로 지정한 정확한 거리로 다시 계산하라는 인스트럭션입니다. k개의 근접 변수를 가져올 때 k*k_factor개의 점에 대해 재계산이 이루어집니다.
+
+# Embedding 테크닉
+## Alpha 쿼리 확장
+퀴리 확장이란 탐색에서 사용되는 기술로, 예를 들어 전문 탐색 시, 입력된 검색문에 단어를 몇 개를 추가함으로써 검색 정확도를 올리는 방법입니다. 백터 탐색을 위해서도 몇가지 방법이 제안되었는데, 그 중 α-쿼리 확장은 추가 학습이 필요 없는 매우 효과적인 방법으로 알려져 있습니다. [Attention-Based Query Expansion Learning](https://arxiv.org/abs/2007.08019)와 [2nd place solution of kaggle shopee competition](https://www.kaggle.com/code/lyakaap/2nd-place-solution/notebook) 논문에서 소개된 바 있습니다..
+
+α-쿼리 확장은 한 벡터에 인접한 벡터를 유사도의 α곱한 가중치로 더해주면 됩니다. 코드로 예시를 들어 보겠습니다. big_npy를 α query expansion로 대체합니다.
+
+```python
+alpha = 3.
+index = Faiss.index_factory(256, "IVF512,PQ128x4fs,RFlat")
+original_norm = np.maximum(np.linalg.norm(big_npy, ord=2, axis=1, keepdims=True), 1e-9)
+big_npy /= original_norm
+index.train(big_npy)
+index.add(big_npy)
+dist, neighbor = index.search(big_npy, num_expand)
+
+expand_arrays = []
+ixs = np.arange(big_npy.shape[0])
+for i in range(-(-big_npy.shape[0]//batch_size)):
+ ix = ixs[i*batch_size:(i+1)*batch_size]
+ weight = np.power(np.einsum("nd,nmd->nm", big_npy[ix], big_npy[neighbor[ix]]), alpha)
+ expand_arrays.append(np.sum(big_npy[neighbor[ix]] * np.expand_dims(weight, axis=2),axis=1))
+big_npy = np.concatenate(expand_arrays, axis=0)
+
+# index version 정규화
+big_npy = big_npy / np.maximum(np.linalg.norm(big_npy, ord=2, axis=1, keepdims=True), 1e-9)
+```
+
+위 테크닉은 탐색을 수행하는 쿼리에도, 탐색 대상 DB에도 적응 가능한 테크닉입니다.
+
+## MiniBatch KMeans에 의한 embedding 압축
+
+total_fea.npy가 너무 클 경우 K-means를 이용하여 벡터를 작게 만드는 것이 가능합니다. 이하 코드로 embedding의 압축이 가능합니다. n_clusters에 압축하고자 하는 크기를 지정하고 batch_size에 256 * CPU의 코어 수를 지정함으로써 CPU 병렬화의 혜택을 충분히 얻을 수 있습니다.
+
+```python
+import multiprocessing
+from sklearn.cluster import MiniBatchKMeans
+kmeans = MiniBatchKMeans(n_clusters=10000, batch_size=256 * multiprocessing.cpu_count(), init="random")
+kmeans.fit(big_npy)
+sample_npy = kmeans.cluster_centers_
+```
\ No newline at end of file
diff --git a/docs/kr/faq_ko.md b/docs/kr/faq_ko.md
new file mode 100644
index 0000000000000000000000000000000000000000..b99a0813f56d7b95d04cbbe7d8d2b3017fbbfbea
--- /dev/null
+++ b/docs/kr/faq_ko.md
@@ -0,0 +1,125 @@
+## Q1:일괄 트레이닝이 끝나고 인덱스가 없음
+
+"Training is done. The program is closed."라고 표시되면 모델 트레이닝이 성공한 것이며, 이어지는 오류는 가짜입니다.
+
+일괄 트레이닝이 끝나고 'added'로 시작하는 인덱스 파일이 없으면 트레이닝 세트가 너무 커서 인덱스 추가 단계에서 멈췄을 수 있습니다. 메모리에 대한 인덱스 추가 요구 사항이 너무 큰 문제를 배치 처리 add 인덱스로 해결했습니다. 임시로 "트레이닝 인덱스" 버튼을 다시 클릭해 보세요.
+
+## Q2:트레이닝이 끝나고 트레이닝 세트의 음색을 추론에서 보지 못함
+
+'음색 새로고침'을 클릭해 보세요. 여전히 없다면 트레이닝에 오류가 있는지, 콘솔 및 webui의 스크린샷, logs/실험명 아래의 로그를 개발자에게 보내 확인해 보세요.
+
+## Q3:모델 공유 방법
+
+rvc_root/logs/실험명 아래에 저장된 pth는 추론에 사용하기 위한 것이 아니라 실험 상태를 저장하고 복원하며, 트레이닝을 계속하기 위한 것입니다. 공유에 사용되는 모델은 weights 폴더 아래 60MB 이상인 pth 파일입니다.
+
+향후에는 weights/exp_name.pth와 logs/exp_name/added_xxx.index를 결합하여 weights/exp_name.zip으로 만들어 index 입력 단계를 생략할 예정입니다. 그러면 zip 파일을 공유하고 pth 파일은 공유하지 마세요. 단지 다른 기계에서 트레이닝을 계속하려는 경우에만 공유하세요.
+
+logs 폴더 아래 수백 MB의 pth 파일을 weights 폴더에 복사/공유하여 강제로 추론에 사용하면 f0, tgt_sr 등의 키가 없다는 오류가 발생할 수 있습니다. ckpt 탭 아래에서 수동 또는 자동(로컬 logs에서 관련 정보를 찾을 수 있는 경우 자동)으로 음성, 대상 오디오 샘플링률 옵션을 선택한 후 ckpt 소형 모델을 추출해야 합니다(입력 경로에 G로 시작하는 경로를 입력). 추출 후 weights 폴더에 60MB 이상의 pth 파일이 생성되며, 음색 새로고침 후 사용할 수 있습니다.
+
+## Q4:연결 오류
+
+아마도 컨트롤 콘솔(검은 창)을 닫았을 것입니다.
+
+## Q5:WebUI에서 "Expecting value: line 1 column 1 (char 0)" 오류가 발생함
+
+시스템 로컬 네트워크 프록시/글로벌 프록시를 닫으세요.
+
+이는 클라이언트의 프록시뿐만 아니라 서버 측의 프록시도 포함합니다(예: autodl로 http_proxy 및 https_proxy를 설정한 경우 사용 시 unset으로 끄세요).
+
+## Q6:WebUI 없이 명령으로 트레이닝 및 추론하는 방법
+
+트레이닝 스크립트:
+먼저 WebUI를 실행하여 데이터 세트 처리 및 트레이닝에 사용되는 명령줄을 메시지 창에서 확인할 수 있습니다.
+
+추론 스크립트:
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+예제:
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#harvest 또는 pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+## Q7:Cuda 오류/Cuda 메모리 부족
+
+아마도 cuda 설정 문제이거나 장치가 지원되지 않을 수 있습니다. 대부분의 경우 메모리가 부족합니다(out of memory).
+
+트레이닝의 경우 batch size를 줄이세요(1로 줄여도 부족하다면 다른 그래픽 카드로 트레이닝을 해야 합니다). 추론의 경우 config.py 파일 끝에 있는 x_pad, x_query, x_center, x_max를 적절히 줄이세요. 4GB 미만의 메모리(예: 1060(3GB) 및 여러 2GB 그래픽 카드)를 가진 경우는 포기하세요. 4GB 메모리 그래픽 카드는 아직 구할 수 있습니다.
+
+## Q8:total_epoch를 몇으로 설정하는 것이 좋을까요
+
+트레이닝 세트의 오디오 품질이 낮고 배경 소음이 많으면 20~30이면 충분합니다. 너무 높게 설정하면 바닥 모델의 오디오 품질이 낮은 트레이닝 세트를 높일 수 없습니다.
+트레이닝 세트의 오디오 품질이 높고 배경 소음이 적고 길이가 길 경우 높게 설정할 수 있습니다. 200도 괜찮습니다(트레이닝 속도가 빠르므로, 고품질 트레이닝 세트를 준비할 수 있는 조건이 있다면, 그래픽 카드도 좋을 것이므로, 조금 더 긴 트레이닝 시간에 대해 걱정하지 않을 것입니다).
+
+## Q9: 트레이닝 세트는 얼마나 길어야 하나요
+
+10분에서 50분을 추천합니다.
+
+음질이 좋고 백그라운드 노이즈가 낮은 상태에서, 개인적인 특색 있는 음색이라면 더 많으면 더 좋습니다.
+
+고품질의 트레이닝 세트(정교하게 준비된 + 특색 있는 음색)라면, 5분에서 10분도 괜찮습니다. 저장소의 저자도 종종 이렇게 합니다.
+
+1분에서 2분의 데이터로 트레이닝에 성공한 사람도 있지만, 그러한 성공 사례는 다른 사람이 재현하기 어려우며 참고 가치가 크지 않습니다. 이는 트레이닝 세트의 음색이 매우 뚜렷해야 하며(예: 높은 주파수의 명확한 목소리나 소녀음) 음질이 좋아야 합니다.
+
+1분 미만의 데이터로 트레이닝을 시도(성공)한 사례는 아직 보지 못했습니다. 이런 시도는 권장하지 않습니다.
+
+## Q10: index rate는 무엇이며, 어떻게 조정하나요? (과학적 설명)
+
+만약 베이스 모델과 추론 소스의 음질이 트레이닝 세트보다 높다면, 그들은 추론 결과의 음질을 높일 수 있지만, 음색이 베이스 모델/추론 소스의 음색으로 기울어질 수 있습니다. 이 현상을 "음색 유출"이라고 합니다.
+
+index rate는 음색 유출 문제를 줄이거나 해결하는 데 사용됩니다. 1로 조정하면 이론적으로 추론 소스의 음색 유출 문제가 없지만, 음질은 트레이닝 세트에 더 가깝게 됩니다. 만약 트레이닝 세트의 음질이 추론 소스보다 낮다면, index rate를 높이면 음질이 낮아질 수 있습니다. 0으로 조정하면 검색 혼합을 이용하여 트레이닝 세트의 음색을 보호하는 효과가 없습니다.
+
+트레이닝 세트가 고품질이고 길이가 길 경우, total_epoch를 높일 수 있으며, 이 경우 모델 자체가 추론 소스와 베이스 모델의 음색을 거의 참조하지 않아 "음색 유출" 문제가 거의 발생하지 않습니다. 이때 index rate는 중요하지 않으며, 심지어 index 색인 파일을 생성하거나 공유하지 않아도 됩니다.
+
+## Q11: 추론시 GPU를 어떻게 선택하나요?
+
+config.py 파일에서 device cuda: 다음에 카드 번호를 선택합니다.
+카드 번호와 그래픽 카드의 매핑 관계는 트레이닝 탭의 그래픽 카드 정보란에서 볼 수 있습니다.
+
+## Q12: 트레이닝 중간에 저장된 pth를 어떻게 추론하나요?
+
+ckpt 탭 하단에서 소형 모델을 추출합니다.
+
+## Q13: 트레이닝을 어떻게 중단하고 계속할 수 있나요?
+
+현재 단계에서는 WebUI 콘솔을 닫고 go-web.bat을 더블 클릭하여 프로그램을 다시 시작해야 합니다. 웹 페이지 매개변수도 새로 고쳐서 다시 입력해야 합니다.
+트레이닝을 계속하려면: 같은 웹 페이지 매개변수로 트레이닝 모델을 클릭하면 이전 체크포인트에서 트레이닝을 계속합니다.
+
+## Q14: 트레이닝 중 파일 페이지/메모리 오류가 발생하면 어떻게 해야 하나요?
+
+프로세스가 너무 많이 열려 메모리가 폭발했습니다. 다음과 같은 방법으로 해결할 수 있습니다.
+
+1. "음높이 추출 및 데이터 처리에 사용되는 CPU 프로세스 수"를 적당히 낮춥니다.
+2. 트레이닝 세트 오디오를 수동으로 잘라 너무 길지 않게 합니다.
+
+## Q15: 트레이닝 도중 데이터를 어떻게 추가하나요?
+
+1. 모든 데이터에 새로운 실험 이름을 만듭니다.
+2. 이전에 가장 최신의 G와 D 파일(또는 어떤 중간 ckpt를 기반으로 트레이닝하고 싶다면 중간 것을 복사할 수도 있음)을 새 실험 이름으로 복사합니다.
+3. 새 실험 이름으로 원클릭 트레이닝을 시작하면 이전의 최신 진행 상황에서 계속 트레이닝합니다.
+
+## Q16: llvmlite.dll에 관한 오류
+
+```bash
+OSError: Could not load shared object file: llvmlite.dll
+
+FileNotFoundError: Could not find module lib\site-packages\llvmlite\binding\llvmlite.dll (or one of its dependencies). Try using the full path with constructor syntax.
+```
+
+Windows 플랫폼에서 이 오류가 발생하면 https://aka.ms/vs/17/release/vc_redist.x64.exe를 설치하고 WebUI를 다시 시작하면 해결됩니다.
+
+## Q17: RuntimeError: 텐서의 확장된 크기(17280)는 비 단일 항목 차원 1에서 기존 크기(0)와 일치해야 합니다. 대상 크기: [1, 17280]. 텐서 크기: [0]
+
+wavs16k 폴더 아래에서 다른 파일들보다 크기가 현저히 작은 일부 오디오 파일을 찾아 삭제하고, 트레이닝 모델을 클릭하면 오류가 발생하지 않습니다. 하지만 원클릭 프로세스가 중단되었기 때문에 모델 트레이닝이 완료된 후에는 인덱스 트레이닝을 클릭해야 합니다.
+
+## Q18: RuntimeError: 텐서 a의 크기(24)가 비 단일 항목 차원 2에서 텐서 b(16)의 크기와 일치해야 합니다.
+
+트레이닝 도중에 샘플링 레이트를 변경해서는 안 됩니다. 변경해야 한다면 실험 이름을 변경하고 처음부터 트레이닝해야 합니다. 물론, 이전에 추출한 음높이와 특징(0/1/2/2b 폴더)을 복사하여 트레이닝 프로세스를 가속화할 수도 있습니다.
diff --git a/docs/kr/training_tips_ko.md b/docs/kr/training_tips_ko.md
new file mode 100644
index 0000000000000000000000000000000000000000..8b3b6245862aef69480f57263d268c94d5e843ca
--- /dev/null
+++ b/docs/kr/training_tips_ko.md
@@ -0,0 +1,53 @@
+RVC 훈련에 대한 설명과 팁들
+======================================
+본 팁에서는 어떻게 데이터 훈련이 이루어지고 있는지 설명합니다.
+
+# 훈련의 흐름
+GUI의 훈련 탭의 단계를 따라 설명합니다.
+
+## step1
+실험 이름을 지정합니다. 또한, 모델이 피치(소리의 높낮이)를 고려해야 하는지 여부를 여기에서 설정할 수도 있습니다..
+각 실험을 위한 데이터는 `/logs/experiment name/`에 배치됩니다..
+
+## step2a
+음성 파일을 불러오고 전처리합니다.
+
+### 음성 파일 불러오기
+음성 파일이 있는 폴더를 지정하면 해당 폴더에 있는 음성 파일이 자동으로 가져와집니다.
+예를 들어 `C:Users\hoge\voices`를 지정하면 `C:Users\hoge\voices\voice.mp3`가 읽히지만 `C:Users\hoge\voices\dir\voice.mp3`는 읽히지 않습니다.
+
+음성 로드에는 내부적으로 ffmpeg를 이용하고 있으므로, ffmpeg로 대응하고 있는 확장자라면 자동적으로 읽힙니다.
+ffmpeg에서 int16으로 변환한 후 float32로 변환하고 -1과 1 사이에 정규화됩니다.
+
+### 잡음 제거
+음성 파일에 대해 scipy의 filtfilt를 이용하여 잡음을 처리합니다.
+
+### 음성 분할
+입력한 음성 파일은 먼저 일정 기간(max_sil_kept=5초?)보다 길게 무음이 지속되는 부분을 감지하여 음성을 분할합니다.무음으로 음성을 분할한 후에는 0.3초의 overlap을 포함하여 4초마다 음성을 분할합니다.4초 이내에 구분된 음성은 음량의 정규화를 실시한 후 wav 파일을 `/logs/실험명/0_gt_wavs`로, 거기에서 16k의 샘플링 레이트로 변환해 `/logs/실험명/1_16k_wavs`에 wav 파일로 저장합니다.
+
+## step2b
+### 피치 추출
+wav 파일에서 피치(소리의 높낮이) 정보를 추출합니다. parselmouth나 pyworld에 내장되어 있는 메서드으로 피치 정보(=f0)를 추출해, `/logs/실험명/2a_f0`에 저장합니다. 그 후 피치 정보를 로그로 변환하여 1~255 정수로 변환하고 `/logs/실험명/2b-f0nsf`에 저장합니다.
+
+### feature_print 추출
+HuBERT를 이용하여 wav 파일을 미리 embedding으로 변환합니다. `/logs/실험명/1_16k_wavs`에 저장한 wav 파일을 읽고 HuBERT에서 wav 파일을 256차원 feature들로 변환한 후 npy 형식으로 `/logs/실험명/3_feature256`에 저장합니다.
+
+## step3
+모델의 훈련을 진행합니다.
+
+### 초보자용 용어 해설
+심층학습(딥러닝)에서는 데이터셋을 분할하여 조금씩 학습을 진행합니다.한 번의 모델 업데이트(step) 단계 당 batch_size개의 데이터를 탐색하여 예측과 오차를 수정합니다. 데이터셋 전부에 대해 이 작업을 한 번 수행하는 이를 하나의 epoch라고 계산합니다.
+
+따라서 학습 시간은 단계당 학습 시간 x (데이터셋 내 데이터의 수 / batch size) x epoch 수가 소요됩니다. 일반적으로 batch size가 클수록 학습이 안정적이게 됩니다. (step당 학습 시간 ÷ batch size)는 작아지지만 GPU 메모리를 더 많이 사용합니다. GPU RAM은 nvidia-smi 명령어를 통해 확인할 수 있습니다. 실행 환경에 따라 배치 크기를 최대한 늘리면 짧은 시간 내에 학습이 가능합니다.
+
+### 사전 학습된 모델 지정
+RVC는 적은 데이터셋으로도 훈련이 가능하도록 사전 훈련된 가중치에서 모델 훈련을 시작합니다. 기본적으로 `rvc-location/pretrained/f0G40k.pth` 및 `rvc-location/pretrained/f0D40k.pth`를 불러옵니다. 학습을 할 시에, 모델 파라미터는 각 save_every_epoch별로 `logs/experiment name/G_{}.pth` 와 `logs/experiment name/D_{}.pth`로 저장이 되는데, 이 경로를 지정함으로써 학습을 재개하거나, 다른 실험에서 학습한 모델의 가중치에서 학습을 시작할 수 있습니다.
+
+### index의 학습
+RVC에서는 학습시에 사용된 HuBERT의 feature값을 저장하고, 추론 시에는 학습 시 사용한 feature값과 유사한 feature 값을 탐색해 추론을 진행합니다. 이 탐색을 고속으로 수행하기 위해 사전에 index을 학습하게 됩니다.
+Index 학습에는 근사 근접 탐색법 라이브러리인 Faiss를 사용하게 됩니다. `/logs/실험명/3_feature256`의 feature값을 불러와, 이를 모두 결합시킨 feature값을 `/logs/실험명/total_fea.npy`로서 저장, 그것을 사용해 학습한 index를`/logs/실험명/add_XXX.index`로 저장합니다.
+
+### 버튼 설명
+- モデルのトレーニング (모델 학습): step2b까지 실행한 후, 이 버튼을 눌러 모델을 학습합니다.
+- 特徴インデックスのトレーニング (특징 지수 훈련): 모델의 훈련 후, index를 학습합니다.
+- ワンクリックトレーニング (원클릭 트레이닝): step2b까지의 모델 훈련, feature index 훈련을 일괄로 실시합니다.
\ No newline at end of file
diff --git a/docs/pt/README.pt.md b/docs/pt/README.pt.md
new file mode 100644
index 0000000000000000000000000000000000000000..4d66a5e3349e568d7cd1c5235528f6e61310041b
--- /dev/null
+++ b/docs/pt/README.pt.md
@@ -0,0 +1,180 @@
+
+
+# Retrieval-based-Voice-Conversion-WebUI
+Uma estrutura de conversão de voz fácil de usar baseada em VITS.
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+
+
+------
+[**FAQ (Frequently Asked Questions)**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/FAQ-(Frequently-Asked-Questions))
+
+[**English**](../en/README.en.md) | [**中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+
+Confira nosso [Vídeo de demonstração](https://www.bilibili.com/video/BV1pm4y1z7Gm/) aqui!
+
+Treinamento/Inferência WebUI：go-web.bat
+
+
+GUI de conversão de voz em tempo real：go-realtime-gui.bat
+
+
+
+> O dataset para o modelo de pré-treinamento usa quase 50 horas de conjunto de dados de código aberto VCTK de alta qualidade.
+
+> Dataset de músicas licenciadas de alta qualidade serão adicionados ao conjunto de treinamento, um após o outro, para seu uso, sem se preocupar com violação de direitos autorais.
+
+> Aguarde o modelo básico pré-treinado do RVCv3, que possui parâmetros maiores, mais dados de treinamento, melhores resultados, velocidade de inferência inalterada e requer menos dados de treinamento para treinamento.
+
+## Resumo
+Este repositório possui os seguintes recursos:
++ Reduza o vazamento de tom substituindo o recurso de origem pelo recurso de conjunto de treinamento usando a recuperação top1;
++ Treinamento fácil e rápido, mesmo em placas gráficas relativamente ruins;
++ Treinar com uma pequena quantidade de dados também obtém resultados relativamente bons (>=10min de áudio com baixo ruído recomendado);
++ Suporta fusão de modelos para alterar timbres (usando guia de processamento ckpt-> mesclagem ckpt);
++ Interface Webui fácil de usar;
++ Use o modelo UVR5 para separar rapidamente vocais e instrumentos.
++ Use o mais poderoso algoritmo de extração de voz de alta frequência [InterSpeech2023-RMVPE](#Credits) para evitar o problema de som mudo. Fornece os melhores resultados (significativamente) e é mais rápido, com consumo de recursos ainda menor que o Crepe_full.
++ Suporta aceleração de placas gráficas AMD/Intel.
++ Aceleração de placas gráficas Intel ARC com suporte para IPEX.
+
+## Preparando o ambiente
+Os comandos a seguir precisam ser executados no ambiente Python versão 3.8 ou superior.
+
+(Windows/Linux)
+Primeiro instale as dependências principais através do pip:
+```bash
+# Instale as dependências principais relacionadas ao PyTorch, pule se instaladas
+# Referência: https://pytorch.org/get-started/locally/
+pip install torch torchvision torchaudio
+
+#Para arquitetura Windows + Nvidia Ampere (RTX30xx), você precisa especificar a versão cuda correspondente ao pytorch de acordo com a experiência de https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/issues/ 21
+#pip instalar tocha torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+
+#Para placas Linux + AMD, você precisa usar as seguintes versões do pytorch:
+#pip instalar tocha torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm5.4.2
+```
+
+Então pode usar poesia para instalar as outras dependências:
+```bash
+# Instale a ferramenta de gerenciamento de dependências Poetry, pule se instalada
+# Referência: https://python-poetry.org/docs/#installation
+curl -sSL https://install.python-poetry.org | python3 -
+
+#Instale as dependências do projeto
+poetry install
+```
+
+Você também pode usar pip para instalá-los:
+```bash
+
+for Nvidia graphics cards
+ pip install -r requirements/main.txt
+
+for AMD/Intel graphics cards on Windows (DirectML)：
+ pip install -r requirements/dml.txt
+
+for Intel ARC graphics cards on Linux / WSL using Python 3.10:
+ pip install -r requirements/ipex.txt
+
+for AMD graphics cards on Linux (ROCm):
+ pip install -r requirements/amd.txt
+```
+
+------
+Usuários de Mac podem instalar dependências via `run.sh`:
+```bash
+sh ./run.sh
+```
+
+## Preparação de outros Pré-modelos
+RVC requer outros pré-modelos para inferir e treinar.
+
+```bash
+#Baixe todos os modelos necessários em https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/
+python tools/download_models.py
+```
+
+Ou apenas baixe-os você mesmo em nosso [Huggingface space](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/).
+
+Aqui está uma lista de pré-modelos e outros arquivos que o RVC precisa:
+```bash
+./assets/hubert/hubert_base.pt
+
+./assets/pretrained
+
+./assets/uvr5_weights
+
+Downloads adicionais são necessários se você quiser testar a versão v2 do modelo.
+
+./assets/pretrained_v2
+
+Se você deseja testar o modelo da versão v2 (o modelo da versão v2 alterou a entrada do recurso dimensional 256 do Hubert + final_proj de 9 camadas para o recurso dimensional 768 do Hubert de 12 camadas e adicionou 3 discriminadores de período), você precisará baixar recursos adicionais
+
+./assets/pretrained_v2
+
+Se quiser usar o algoritmo de extração de tom vocal SOTA RMVPE mais recente, você precisa baixar os pesos RMVPE e colocá-los no diretório raiz RVC
+
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt
+
+ Para usuários de placas gráficas AMD/Intel, você precisa baixar:
+
+ https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx
+
+```
+
+Os usuários de placas gráficas Intel ARC precisam executar o comando `source /opt/intel/oneapi/setvars.sh` antes de iniciar o Webui.
+
+Em seguida, use este comando para iniciar o Webui:
+```bash
+python web.py
+```
+
+Se estiver usando Windows ou macOS, você pode baixar e extrair `RVC-beta.7z` para usar RVC diretamente usando `go-web.bat` no Windows ou `sh ./run.sh` no macOS para iniciar o Webui.
+
+## Suporte ROCm para placas gráficas AMD (somente Linux)
+Para usar o ROCm no Linux, instale todos os drivers necessários conforme descrito [aqui](https://rocm.docs.amd.com/en/latest/deploy/linux/os-native/install.html).
+
+No Arch use pacman para instalar o driver:
+````
+pacman -S rocm-hip-sdk rocm-opencl-sdk
+````
+
+Talvez você também precise definir estas variáveis de ambiente (por exemplo, em um RX6700XT):
+````
+export ROCM_PATH=/opt/rocm
+export HSA_OVERRIDE_GFX_VERSION=10.3.0
+````
+Verifique também se seu usuário faz parte do grupo `render` e `video`:
+````
+sudo usermod -aG render $USERNAME
+sudo usermod -aG video $USERNAME
+````
+Depois disso, você pode executar o WebUI:
+```bash
+python web.py
+```
+
+## Credits
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
++ [Vocal pitch extraction:RMVPE](https://github.com/Dream-High/RMVPE)
+ + The pretrained model is trained and tested by [yxlllc](https://github.com/yxlllc/RMVPE) and [RVC-Boss](https://github.com/RVC-Boss).
+
+## Thanks to all contributors for their efforts
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
+
diff --git a/docs/pt/faiss_tips_pt.md b/docs/pt/faiss_tips_pt.md
new file mode 100644
index 0000000000000000000000000000000000000000..8649979b599642ec54306c78aace6478130a1250
--- /dev/null
+++ b/docs/pt/faiss_tips_pt.md
@@ -0,0 +1,102 @@
+pONTAS de afinação FAISS
+==================
+# sobre faiss
+faiss é uma biblioteca de pesquisas de vetores densos na área, desenvolvida pela pesquisa do facebook, que implementa com eficiência muitos métodos de pesquisa de área aproximada.
+A Pesquisa Aproximada de área encontra vetores semelhantes rapidamente, sacrificando alguma precisão.
+
+## faiss em RVC
+No RVC, para a incorporação de recursos convertidos pelo HuBERT, buscamos incorporações semelhantes à incorporação gerada a partir dos dados de treinamento e as misturamos para obter uma conversão mais próxima do discurso original. No entanto, como essa pesquisa leva tempo se realizada de forma ingênua, a conversão de alta velocidade é realizada usando a pesquisa aproximada de área.
+
+# visão geral da implementação
+Em '/logs/nome-do-seu-modelo/3_feature256', onde o modelo está localizado, os recursos extraídos pelo HuBERT de cada dado de voz estão localizados.
+A partir daqui, lemos os arquivos npy ordenados por nome de arquivo e concatenamos os vetores para criar big_npy. (Este vetor tem a forma [N, 256].)
+Depois de salvar big_npy as /logs/nome-do-seu-modelo/total_fea.npy, treine-o com faiss.
+
+Neste artigo, explicarei o significado desses parâmetros.
+
+# Explicação do método
+## Fábrica de Index
+Uma fábrica de Index é uma notação faiss exclusiva que expressa um pipeline que conecta vários métodos de pesquisa de área aproximados como uma string.
+Isso permite que você experimente vários métodos aproximados de pesquisa de área simplesmente alterando a cadeia de caracteres de fábrica do Index.
+No RVC é usado assim:
+
+```python
+index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+Entre os argumentos de index_factory, o primeiro é o número de dimensões do vetor, o segundo é a string de fábrica do Index e o terceiro é a distância a ser usada.
+
+Para uma notação mais detalhada
+https://github.com/facebookresearch/faiss/wiki/The-index-factory
+
+## Construção de Index
+Existem dois Indexs típicos usados como similaridade de incorporação da seguinte forma.
+
+- Distância euclidiana (MÉTRICA_L2)
+- Produto interno (METRIC_INNER_PRODUCT)
+
+A distância euclidiana toma a diferença quadrática em cada dimensão, soma as diferenças em todas as dimensões e, em seguida, toma a raiz quadrada. Isso é o mesmo que a distância em 2D e 3D que usamos diariamente.
+O produto interno não é usado como um Index de similaridade como é, e a similaridade de cosseno que leva o produto interno depois de ser normalizado pela norma L2 é geralmente usada.
+
+O que é melhor depende do caso, mas a similaridade de cosseno é frequentemente usada na incorporação obtida pelo word2vec e modelos de recuperação de imagem semelhantes aprendidos pelo ArcFace. Se você quiser fazer a normalização l2 no vetor X com numpy, você pode fazê-lo com o seguinte código com eps pequeno o suficiente para evitar a divisão 0.
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+Além disso, para a Construção de Index, você pode alterar o Index de distância usado para cálculo escolhendo o valor a ser passado como o terceiro argumento.
+
+```python
+index = faiss.index_factory(dimention, text, faiss.METRIC_INNER_PRODUCT)
+```
+
+## FI
+IVF (Inverted file indexes) é um algoritmo semelhante ao Index invertido na pesquisa de texto completo.
+Durante o aprendizado, o destino da pesquisa é agrupado com kmeans e o particionamento Voronoi é realizado usando o centro de cluster. A cada ponto de dados é atribuído um cluster, por isso criamos um dicionário que procura os pontos de dados dos clusters.
+
+Por exemplo, se os clusters forem atribuídos da seguinte forma
+|index|Cluster|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+O Index invertido resultante se parece com isso:
+
+| cluster | Index |
+|-------|-----|
+| A | 1, 3 |
+| B | 2 5 |
+| C | 4 |
+
+Ao pesquisar, primeiro pesquisamos n_probe clusters dos clusters e, em seguida, calculamos as distâncias para os pontos de dados pertencentes a cada cluster.
+
+# Parâmetro de recomendação
+Existem diretrizes oficiais sobre como escolher um Index, então vou explicar de
+acordo. https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
+
+Para conjuntos de dados abaixo de 1M, o 4bit-PQ é o método mais eficiente disponível no faiss em abril de 2023.
+Combinando isso com a fertilização in vitro, estreitando os candidatos com 4bit-PQ e, finalmente, recalcular a distância com um Index preciso pode ser descrito usando a seguinte fábrica de Indexs.
+
+```python
+index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## Parâmetros recomendados para FIV
+Considere o caso de muitas FIVs. Por exemplo, se a quantização grosseira por FIV for realizada para o número de dados, isso é o mesmo que uma pesquisa exaustiva ingênua e é ineficiente.
+Para 1M ou menos, os valores de FIV são recomendados entre 4*sqrt(N) ~ 16*sqrt(N) para N número de pontos de dados.
+
+Como o tempo de cálculo aumenta proporcionalmente ao número de n_sondas, consulte a precisão e escolha adequadamente. Pessoalmente, não acho que o RVC precise de tanta precisão, então n_probe = 1 está bem.
+
+## FastScan
+O FastScan é um método que permite a aproximação de alta velocidade de distâncias por quantização de produto cartesiano, realizando-as em registros.
+A quantização cartesiana do produto executa o agrupamento independentemente para cada dimensão d (geralmente d = 2) durante o aprendizado, calcula a distância entre os agrupamentos com antecedência e cria uma tabela de pesquisa. No momento da previsão, a distância de cada dimensão pode ser calculada em O(1) olhando para a tabela de pesquisa.
+Portanto, o número que você especifica após PQ geralmente especifica metade da dimensão do vetor.
+
+Para uma descrição mais detalhada do FastScan, consulte a documentação oficial.
+https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlat é uma instrução para recalcular a distância aproximada calculada pelo FastScan com a distância exata especificada pelo terceiro argumento da Construção de Index.
+Ao obter áreas k, os pontos k*k_factor são recalculados.
diff --git a/docs/pt/faq_pt.md b/docs/pt/faq_pt.md
new file mode 100644
index 0000000000000000000000000000000000000000..6a2a3de42bc837d1c11186cd74b8e40827a2b376
--- /dev/null
+++ b/docs/pt/faq_pt.md
@@ -0,0 +1,220 @@
+# FAQ AI HUB BRASIL
+## O que é epoch, quantos utilizar, quanto de dataset utilizar e qual à configuração interessante?
+Epochs basicamente quantas vezes o seu dataset foi treinado.
+
+Recomendado ler Q8 e Q9 no final dessa página pra entender mais sobre dataset e epochs
+
+__**Não é uma regra, mas opinião:**__
+
+### **Mangio-Crepe Hop Length**
+- 64 pra cantores e dubladores
+- 128(padrão) para os demais (editado)
+
+### **Epochs e dataset**
+600epoch para cantores - --dataset entre 10 e 50 min desnecessario mais que 50 minutos--
+300epoch para os demais - --dataset entre 10 e 50 min desnecessario mais que 50 minutos--
+
+### **Tom**
+magio-crepe se for audios extraído de alguma musica
+harvest se for de estúdio
+
+## O que é index?
+Basicamente o que define o sotaque. Quanto maior o numero, mas próximo o sotaque fica do original. Porém, quando o modelo é bem, não é necessário um index.
+
+## O que significa cada sigla (pm, harvest, crepe, magio-crepe, RMVPE)?
+
+- pm = extração mais rápida, mas discurso de qualidade inferior;
+- harvest = graves melhores, mas extremamente lentos;
+- dio = conversão rápida mas pitch ruim;
+- crepe = melhor qualidade, mas intensivo em GPU;
+- crepe-tiny = mesma coisa que o crepe, só que com a qualidade um pouco inferior;
+- **mangio-crepe = melhor qualidade, mais otimizado; (MELHOR OPÇÃO)**
+- mangio-crepe-tiny = mesma coisa que o mangio-crepe, só que com a qualidade um pouco inferior;
+- RMVPE: um modelo robusto para estimativa de afinação vocal em música polifônica;
+
+## Pra rodar localmente, quais os requisitos minimos?
+Já tivemos relatos de pessoas com GTX 1050 rodando inferencia, se for treinar numa 1050 vai demorar muito mesmo e inferior a isso, normalmente da tela azul
+
+O mais importante é placa de vídeo, vram na verdade
+Se você tiver 4GB ou mais, você tem uma chance.
+
+**NOS DOIS CASOS NÃO É RECOMENDADO UTILIZAR O PC ENQUANTO ESTÁ UTILIZNDO, CHANCE DE TELA AZUL É ALTA**
+### Inference
+Não é algo oficial para requisitos minimos
+- Placa de vídeo: nvidia de 4gb
+- Memoria ram: 8gb
+- CPU: ?
+- Armanezamento: 20gb (sem modelos)
+
+### Treinamento de voz
+Não é algo oficial para requisitos minimos
+- Placa de vídeo: nvidia de 6gb
+- Memoria ram: 16gb
+- CPU: ?
+- Armanezamento: 20gb (sem modelos)
+
+## Limite de GPU no Google Colab excedido, apenas CPU o que fazer?
+Recomendamos esperar outro dia pra liberar mais 15gb ou 12 horas pra você. Ou você pode contribuir com o Google pagando algum dos planos, ai aumenta seu limite.
+Utilizar apenas CPU no Google Colab demora DEMAIS.
+
+
+## Google Colab desconectando com muita frequencia, o que fazer?
+Neste caso realmente não tem muito o que fazer. Apenas aguardar o proprietário do código corrigir ou a gente do AI HUB Brasil achar alguma solução. Isso acontece por diversos motivos, um incluindo a Google barrando o treinamento de voz.
+
+## O que é Batch Size/Tamanho de lote e qual numero utilizar?
+Batch Size/Tamanho do lote é basicamente quantos epoch faz ao mesmo tempo. Se por 20, ele fazer 20 epoch ao mesmo tempo e isso faz pesar mais na máquina e etc.
+
+No Google Colab você pode utilizar até 20 de boa.
+Se rodando localmente, depende da sua placa de vídeo, começa por baixo (6) e vai testando.
+
+## Sobre backup na hora do treinamento
+Backup vai de cada um. Eu quando uso a ``easierGUI`` utilizo a cada 100 epoch (meu caso isolado).
+No colab, se instavel, coloque a cada 10 epoch
+Recomendo utilizarem entre 25 e 50 pra garantir.
+
+Lembrando que cada arquivo geral é por volta de 50mb, então tenha muito cuidado quanto você coloca. Pois assim pode acabar lotando seu Google Drive ou seu PC.
+
+Depois de finalizado, da pra apagar os epoch de backup.
+
+## Como continuar da onde parou pra fazer mais epochs?
+Primeira coisa que gostaria de lembrar, não necessariamente quanto mais epochs melhor. Se fizer epochs demais vai dar **overtraining** o que pode ser ruim.
+
+### GUI NORMAL
+- Inicie normalmente a GUI novamente.
+- Na aba de treino utilize o MESMO nome que estava treinando, assim vai continuar o treino onde parou o ultimo backup.
+- Ignore as opções ``Processar o Conjunto de dados`` e ``Extrair Tom``
+- Antes de clicar pra treinar, arrume os epoch, bakcup e afins.
+ - Obviamente tem que ser um numero maior do qu estava em epoch.
+ - Backup você pode aumentar ou diminuir
+- Agora você vai ver a opção ``Carregue o caminho G do modelo base pré-treinado:`` e ``Carregue o caminho D do modelo base pré-treinado:``
+ -Aqui você vai por o caminho dos modelos que estão em ``./logs/minha-voz``
+ - Vai ficar algo parecido com isso ``e:/RVC/logs/minha-voz/G_0000.pth`` e ``e:/RVC/logs/minha-voz/D_0000.pth``
+-Coloque pra treinar
+
+**Lembrando que a pasta logs tem que ter todos os arquivos e não somente o arquivo ``G`` e ``D``**
+
+### EasierGUI
+- Inicie normalmente a easierGUI novamente.
+- Na aba de treino utilize o MESMO nome que estava treinando, assim vai continuar o treino onde parou o ultimo backup.
+- Selecione 'Treinar modelo', pode pular os 2 primeiros passos já que vamos continuar o treino.
+
+
+# FAQ Original traduzido
+
+## Q1:Não é possível encontrar o arquivo de Index após "Treinamento com um clique".
+Se exibir "O treinamento está concluído. O programa é fechado ", então o modelo foi treinado com sucesso e os erros subsequentes são falsos;
+
+A falta de um arquivo de index 'adicionado' após o treinamento com um clique pode ser devido ao conjunto de treinamento ser muito grande, fazendo com que a adição do index fique presa; isso foi resolvido usando o processamento em lote para adicionar o index, o que resolve o problema de sobrecarga de memória ao adicionar o index. Como solução temporária, tente clicar no botão "Treinar Index" novamente.
+
+## Q2:Não é possível encontrar o modelo em “Modelo de voz” após o treinamento
+Clique em "Atualizar lista de voz" ou "Atualizar na EasyGUI e verifique novamente; se ainda não estiver visível, verifique se há erros durante o treinamento e envie capturas de tela do console, da interface do usuário da Web e dos ``logs/experiment_name/*.log`` para os desenvolvedores para análise posterior.
+
+## Q3:Como compartilhar um modelo/Como usar os modelos dos outros?
+Os arquivos ``.pth`` armazenados em ``*/logs/minha-voz`` não são destinados para compartilhamento ou inference, mas para armazenar os checkpoits do experimento para reprodutibilidade e treinamento adicional. O modelo a ser compartilhado deve ser o arquivo ``.pth`` de 60+MB na pasta **weights**;
+
+No futuro, ``weights/minha-voz.pth`` e ``logs/minha-voz/added_xxx.index`` serão mesclados em um único arquivo de ``weights/minha-voz.zip`` para eliminar a necessidade de entrada manual de index; portanto, compartilhe o arquivo zip, não somente o arquivo .pth, a menos que você queira continuar treinando em uma máquina diferente;
+
+Copiar/compartilhar os vários arquivos .pth de centenas de MB da pasta de logs para a pasta de weights para inference forçada pode resultar em erros como falta de f0, tgt_sr ou outras chaves. Você precisa usar a guia ckpt na parte inferior para manualmente ou automaticamente (se as informações forem encontradas nos ``logs/minha-voz``), selecione se deseja incluir informações de tom e opções de taxa de amostragem de áudio de destino e, em seguida, extrair o modelo menor. Após a extração, haverá um arquivo pth de 60+ MB na pasta de weights, e você pode atualizar as vozes para usá-lo.
+
+## Q4 Erro de conexão:
+Para sermos otimistas, aperte F5/recarregue a página, pode ter sido apenas um bug da GUI
+
+Se não...
+Você pode ter fechado o console (janela de linha de comando preta).
+Ou o Google Colab, no caso do Colab, as vezes pode simplesmente fechar
+
+## Q5: Pop-up WebUI 'Valor esperado: linha 1 coluna 1 (caractere 0)'.
+Desative o proxy LAN do sistema/proxy global e atualize.
+
+## Q6:Como treinar e inferir sem a WebUI?
+Script de treinamento:
+
Você pode executar o treinamento em WebUI primeiro, e as versões de linha de comando do pré-processamento e treinamento do conjunto de dados serão exibidas na janela de mensagens.
+
+Script de inference:
+
https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+
+por exemplo
+
+``runtime\python.exe myinfer.py 0 "E:\audios\1111.wav" "E:\RVC\logs\minha-voz\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True``
+
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#harvest or pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+## Q7: Erro Cuda/Cuda sem memória.
+Há uma pequena chance de que haja um problema com a configuração do CUDA ou o dispositivo não seja suportado; mais provavelmente, não há memória suficiente (falta de memória).
+
+Para treinamento, reduza o (batch size) tamanho do lote (se reduzir para 1 ainda não for suficiente, talvez seja necessário alterar a placa gráfica); para inference, ajuste as configurações x_pad, x_query, x_center e x_max no arquivo config.py conforme necessário. Cartões de memória 4G ou inferiores (por exemplo, 1060(3G) e várias placas 2G) podem ser abandonados, enquanto os placas de vídeo com memória 4G ainda têm uma chance.
+
+## Q8:Quantos total_epoch são ótimos?
+Se a qualidade de áudio do conjunto de dados de treinamento for ruim e o nível de ruído for alto, **20-30 epochs** são suficientes. Defini-lo muito alto não melhorará a qualidade de áudio do seu conjunto de treinamento de baixa qualidade.
+
+Se a qualidade de áudio do conjunto de treinamento for alta, o nível de ruído for baixo e houver duração suficiente, você poderá aumentá-lo. **200 é aceitável** (uma vez que o treinamento é rápido e, se você puder preparar um conjunto de treinamento de alta qualidade, sua GPU provavelmente poderá lidar com uma duração de treinamento mais longa sem problemas).
+
+## Q9:Quanto tempo de treinamento é necessário?
+
+**Recomenda-se um conjunto de dados de cerca de 10 min a 50 min.**
+
+Com garantia de alta qualidade de som e baixo ruído de fundo, mais pode ser adicionado se o timbre do conjunto de dados for uniforme.
+
+Para um conjunto de treinamento de alto nível (limpo + distintivo), 5min a 10min é bom.
+
+Há algumas pessoas que treinaram com sucesso com dados de 1 a 2 minutos, mas o sucesso não é reproduzível por outros e não é muito informativo.
Isso requer que o conjunto de treinamento tenha um timbre muito distinto (por exemplo, um som de menina de anime arejado de alta frequência) e a qualidade do áudio seja alta;
+Dados com menos de 1 minuto, já obtivemo sucesso. Mas não é recomendado.
+
+
+## Q10:Qual é a taxa do index e como ajustá-la?
+Se a qualidade do tom do modelo pré-treinado e da fonte de inference for maior do que a do conjunto de treinamento, eles podem trazer a qualidade do tom do resultado do inference, mas ao custo de um possível viés de tom em direção ao tom do modelo subjacente/fonte de inference, em vez do tom do conjunto de treinamento, que é geralmente referido como "vazamento de tom".
+
+A taxa de index é usada para reduzir/resolver o problema de vazamento de timbre. Se a taxa do index for definida como 1, teoricamente não há vazamento de timbre da fonte de inference e a qualidade do timbre é mais tendenciosa em relação ao conjunto de treinamento. Se o conjunto de treinamento tiver uma qualidade de som mais baixa do que a fonte de inference, uma taxa de index mais alta poderá reduzir a qualidade do som. Reduzi-lo a 0 não tem o efeito de usar a mistura de recuperação para proteger os tons definidos de treinamento.
+
+Se o conjunto de treinamento tiver boa qualidade de áudio e longa duração, aumente o total_epoch, quando o modelo em si é menos propenso a se referir à fonte inferida e ao modelo subjacente pré-treinado, e há pouco "vazamento de tom", o index_rate não é importante e você pode até não criar/compartilhar o arquivo de index.
+
+## Q11:Como escolher o GPU ao inferir?
+No arquivo ``config.py``, selecione o número da placa em "device cuda:".
+
+O mapeamento entre o número da placa e a placa gráfica pode ser visto na seção de informações da placa gráfica da guia de treinamento.
+
+## Q12:Como usar o modelo salvo no meio do treinamento?
+Salvar via extração de modelo na parte inferior da guia de processamento do ckpt.
+
+## Q13: Erro de arquivo/memória (durante o treinamento)?
+Muitos processos e sua memória não é suficiente. Você pode corrigi-lo por:
+
+1. Diminuir a entrada no campo "Threads da CPU".
+2. Diminuir o tamanho do conjunto de dados.
+
+## Q14: Como continuar treinando usando mais dados
+
+passo 1: coloque todos os dados wav no path2.
+
+etapa 2: exp_name2 + path2 -> processar conjunto de dados e extrair recurso.
+
+passo 3: copie o arquivo G e D mais recente de exp_name1 (seu experimento anterior) para a pasta exp_name2.
+
+passo 4: clique em "treinar o modelo" e ele continuará treinando desde o início da época anterior do modelo exp.
+
+## Q15: erro sobre llvmlite.dll
+
+OSError: Não foi possível carregar o arquivo de objeto compartilhado: llvmlite.dll
+
+FileNotFoundError: Não foi possível encontrar o módulo lib\site-packages\llvmlite\binding\llvmlite.dll (ou uma de suas dependências). Tente usar o caminho completo com sintaxe de construtor.
+
+O problema acontecerá no Windows, instale https://aka.ms/vs/17/release/vc_redist.x64.exe e será corrigido.
+
+## Q16: RuntimeError: O tamanho expandido do tensor (17280) deve corresponder ao tamanho existente (0) na dimensão 1 não singleton. Tamanhos de destino: [1, 17280]. Tamanhos de tensor: [0]
+
+Exclua os arquivos wav cujo tamanho seja significativamente menor que outros e isso não acontecerá novamente. Em seguida, clique em "treinar o modelo" e "treinar o índice".
+
+## Q17: RuntimeError: O tamanho do tensor a (24) deve corresponder ao tamanho do tensor b (16) na dimensão não singleton 2
+
+Não altere a taxa de amostragem e continue o treinamento. Caso seja necessário alterar, o nome do exp deverá ser alterado e o modelo será treinado do zero. Você também pode copiar o pitch e os recursos (pastas 0/1/2/2b) extraídos da última vez para acelerar o processo de treinamento.
+
diff --git a/docs/pt/training_tips_pt.md b/docs/pt/training_tips_pt.md
new file mode 100644
index 0000000000000000000000000000000000000000..985ab4fc1a8af99eb0d968b11a5eba7d3299230b
--- /dev/null
+++ b/docs/pt/training_tips_pt.md
@@ -0,0 +1,65 @@
+Instruções e dicas para treinamento RVC
+======================================
+Estas DICAS explicam como o treinamento de dados é feito.
+
+# Fluxo de treinamento
+Explicarei ao longo das etapas na guia de treinamento da GUI.
+
+## Passo 1
+Defina o nome do experimento aqui.
+
+Você também pode definir aqui se o modelo deve levar em consideração o pitch.
+Se o modelo não considerar o tom, o modelo será mais leve, mas não será adequado para cantar.
+
+Os dados de cada experimento são colocados em `/logs/nome-do-seu-modelo/`.
+
+## Passo 2a
+Carrega e pré-processa áudio.
+
+### Carregar áudio
+Se você especificar uma pasta com áudio, os arquivos de áudio dessa pasta serão lidos automaticamente.
+Por exemplo, se você especificar `C:Users\hoge\voices`, `C:Users\hoge\voices\voice.mp3` será carregado, mas `C:Users\hoge\voices\dir\voice.mp3` será Não carregado.
+
+Como o ffmpeg é usado internamente para leitura de áudio, se a extensão for suportada pelo ffmpeg, ela será lida automaticamente.
+Após converter para int16 com ffmpeg, converta para float32 e normalize entre -1 e 1.
+
+### Eliminar ruído
+O áudio é suavizado pelo filtfilt do scipy.
+
+### Divisão de áudio
+Primeiro, o áudio de entrada é dividido pela detecção de partes de silêncio que duram mais que um determinado período (max_sil_kept=5 segundos?). Após dividir o áudio no silêncio, divida o áudio a cada 4 segundos com uma sobreposição de 0,3 segundos. Para áudio separado em 4 segundos, após normalizar o volume, converta o arquivo wav para `/logs/nome-do-seu-modelo/0_gt_wavs` e, em seguida, converta-o para taxa de amostragem de 16k para `/logs/nome-do-seu-modelo/1_16k_wavs ` como um arquivo wav.
+
+## Passo 2b
+### Extrair pitch
+Extraia informações de pitch de arquivos wav. Extraia as informações de pitch (=f0) usando o método incorporado em Parselmouth ou pyworld e salve-as em `/logs/nome-do-seu-modelo/2a_f0`. Em seguida, converta logaritmicamente as informações de pitch para um número inteiro entre 1 e 255 e salve-as em `/logs/nome-do-seu-modelo/2b-f0nsf`.
+
+### Extrair feature_print
+Converta o arquivo wav para incorporação antecipadamente usando HuBERT. Leia o arquivo wav salvo em `/logs/nome-do-seu-modelo/1_16k_wavs`, converta o arquivo wav em recursos de 256 dimensões com HuBERT e salve no formato npy em `/logs/nome-do-seu-modelo/3_feature256`.
+
+## Passo 3
+treinar o modelo.
+### Glossário para iniciantes
+No aprendizado profundo, o conjunto de dados é dividido e o aprendizado avança aos poucos. Em uma atualização do modelo (etapa), os dados batch_size são recuperados e previsões e correções de erros são realizadas. Fazer isso uma vez para um conjunto de dados conta como um epoch.
+
+Portanto, o tempo de aprendizagem é o tempo de aprendizagem por etapa x (o número de dados no conjunto de dados/tamanho do lote) x o número de epoch. Em geral, quanto maior o tamanho do lote, mais estável se torna o aprendizado (tempo de aprendizado por etapa ÷ tamanho do lote) fica menor, mas usa mais memória GPU. A RAM da GPU pode ser verificada com o comando nvidia-smi. O aprendizado pode ser feito em pouco tempo aumentando o tamanho do lote tanto quanto possível de acordo com a máquina do ambiente de execução.
+
+### Especifique o modelo pré-treinado
+O RVC começa a treinar o modelo a partir de pesos pré-treinados em vez de 0, para que possa ser treinado com um pequeno conjunto de dados.
+
+Por padrão
+
+- Se você considerar o pitch, ele carrega `rvc-location/pretrained/f0G40k.pth` e `rvc-location/pretrained/f0D40k.pth`.
+- Se você não considerar o pitch, ele carrega `rvc-location/pretrained/f0G40k.pth` e `rvc-location/pretrained/f0D40k.pth`.
+
+Ao aprender, os parâmetros do modelo são salvos em `logs/nome-do-seu-modelo/G_{}.pth` e `logs/nome-do-seu-modelo/D_{}.pth` para cada save_every_epoch, mas especificando nesse caminho, você pode começar a aprender. Você pode reiniciar ou iniciar o treinamento a partir de weights de modelo aprendidos em um experimento diferente.
+
+### Index de aprendizado
+O RVC salva os valores de recursos do HuBERT usados durante o treinamento e, durante a inferência, procura valores de recursos que sejam semelhantes aos valores de recursos usados durante o aprendizado para realizar a inferência. Para realizar esta busca em alta velocidade, o index é aprendido previamente.
+Para aprendizagem de index, usamos a biblioteca de pesquisa de associação de áreas aproximadas faiss. Leia o valor do recurso `logs/nome-do-seu-modelo/3_feature256` e use-o para aprender o index, e salve-o como `logs/nome-do-seu-modelo/add_XXX.index`.
+
+(A partir da versão 20230428update, ele é lido do index e não é mais necessário salvar/especificar.)
+
+### Descrição do botão
+- Treinar modelo: Após executar o passo 2b, pressione este botão para treinar o modelo.
+- Treinar índice de recursos: após treinar o modelo, execute o aprendizado do index.
+- Treinamento com um clique: etapa 2b, treinamento de modelo e treinamento de index de recursos, tudo de uma vez.
\ No newline at end of file
diff --git a/docs/tr/README.tr.md b/docs/tr/README.tr.md
new file mode 100644
index 0000000000000000000000000000000000000000..183e7424570c9a20f660aa782877cfeaeecf4495
--- /dev/null
+++ b/docs/tr/README.tr.md
@@ -0,0 +1,141 @@
+
+
+
+# Çekme Temelli Ses Dönüşümü Web Arayüzü
+VITS'e dayalı kullanımı kolay bir Ses Dönüşümü çerçevesi.
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI)
+
+
+
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/blob/main/LICENSE)
+[](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/)
+
+[](https://discord.gg/HcsmBBGyVk)
+
+
+
+------
+[**SSS (Sıkça Sorulan Sorular)**](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/wiki/SSS-(Sıkça-Sorulan-Sorular))
+
+[**İngilizce**](../en/README.en.md) | [**中文简体**](../../README.md) | [**日本語**](../jp/README.ja.md) | [**한국어**](../kr/README.ko.md) ([**韓國語**](../kr/README.ko.han.md)) | [**Français**](../fr/README.fr.md) | [**Türkçe**](../tr/README.tr.md) | [**Português**](../pt/README.pt.md)
+
+Burada [Demo Video'muzu](https://www.bilibili.com/video/BV1pm4y1z7Gm/) izleyebilirsiniz!
+
+RVC Kullanarak Gerçek Zamanlı Ses Dönüşüm Yazılımı: [w-okada/voice-changer](https://github.com/w-okada/voice-changer)
+
+> Ön eğitim modeli için veri kümesi neredeyse 50 saatlik yüksek kaliteli VCTK açık kaynak veri kümesini kullanır.
+
+> Yüksek kaliteli lisanslı şarkı veri setleri telif hakkı ihlali olmadan kullanımınız için eklenecektir.
+
+> Lütfen daha büyük parametrelere, daha fazla eğitim verisine sahip RVCv3'ün ön eğitimli temel modeline göz atın; daha iyi sonuçlar, değişmeyen çıkarsama hızı ve daha az eğitim verisi gerektirir.
+
+## Özet
+Bu depo aşağıdaki özelliklere sahiptir:
++ Ton sızıntısını en aza indirmek için kaynak özelliğini en iyi çıkarımı kullanarak eğitim kümesi özelliği ile değiştirme;
++ Kolay ve hızlı eğitim, hatta nispeten zayıf grafik kartlarında bile;
++ Az miktarda veriyle bile nispeten iyi sonuçlar alın (>=10 dakika düşük gürültülü konuşma önerilir);
++ Timbraları değiştirmek için model birleştirmeyi destekleme (ckpt işleme sekmesi-> ckpt birleştir);
++ Kullanımı kolay Web arayüzü;
++ UVR5 modelini kullanarak hızla vokalleri ve enstrümanları ayırma.
++ En güçlü Yüksek tiz Ses Çıkarma Algoritması [InterSpeech2023-RMVPE](#Krediler) sessiz ses sorununu önlemek için kullanılır. En iyi sonuçları (önemli ölçüde) sağlar ve Crepe_full'den daha hızlı çalışır, hatta daha düşük kaynak tüketimi sağlar.
++ AMD/Intel grafik kartları hızlandırması desteklenir.
++ Intel ARC grafik kartları hızlandırması IPEX ile desteklenir.
+
+## Ortamın Hazırlanması
+Aşağıdaki komutlar, Python sürümü 3.8 veya daha yüksek olan bir ortamda çalıştırılmalıdır.
+
+(Windows/Linux)
+İlk olarak ana bağımlılıkları pip aracılığıyla kurun:
+```bash
+# PyTorch ile ilgili temel bağımlılıkları kurun, zaten kuruluysa atlayın
+# Referans: https://pytorch.org/get-started/locally/
+pip install torch torchvision torchaudio
+
+# Windows + Nvidia Ampere Mimarisi(RTX30xx) için, https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/issues/21 deneyime göre pytorch'a karşılık gelen cuda sürümünü belirtmeniz gerekebilir
+#pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
+```
+
+Sonra poetry kullanarak diğer bağımlılıkları kurabilirsiniz:
+```bash
+# Poetry bağımlılık yönetim aracını kurun, zaten kuruluysa atlayın
+# Referans: https://python-poetry.org/docs/#installation
+curl -sSL https://install.python-poetry.org | python3 -
+
+# Projeyi bağımlılıkları kurun
+poetry install
+```
+
+Ayrıca bunları pip kullanarak da kurabilirsiniz:
+```bash
+
+Nvidia grafik kartları için
+ pip install -r requirements/main.txt
+
+AMD/Intel grafik kartları için：
+ pip install -r requirements/dml.txt
+
+Intel ARC grafik kartları için Linux / WSL ile Python 3.10 kullanarak:
+ pip install -r requirements/ipex.txt
+
+```
+
+------
+Mac kullanıcıları `run.sh` aracılığıyla bağımlılıkları kurabilir:
+```bash
+sh ./run.sh
+```
+
+## Diğer Ön Modellerin Hazırlanması
+RVC'nin çıkarım ve eğitim yapması için diğer ön modellere ihtiyacı vardır.
+
+Bu ön modelleri [Huggingface alanımızdan](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/) indirmeniz gerekecektir.
+
+İşte RVC'nin ihtiyaç duyduğu diğer ön modellerin ve dosyaların bir listesi:
+```bash
+./assets/hubert/hubert_base.pt
+
+./assets/pretrained
+
+./assets/uvr5_weights
+
+V2 sürümü modelini test etmek isterseniz, ek özellikler indirmeniz gerekecektir.
+
+./assets/pretrained_v2
+
+V2 sürüm modelini test etmek isterseniz (v2 sürüm modeli, 9 katmanlı Hubert+final_proj'ün 256 boyutlu özelliğini 12 katmanlı Hubert'ün 768 boyutlu özelliğiyle değiştirmiştir ve 3 periyot ayırıcı eklemiştir), ek özellikleri indirmeniz gerekecektir.
+
+./assets/pretrained_v2
+
+En son SOTA RMVPE vokal ton çıkarma algoritmasını kullanmak istiyorsanız, RMVPE ağırlıklarını indirip RVC kök dizinine koymalısınız.
+
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.pt
+
+ AMD/Intel grafik kartları kullanıcıları için indirmeniz gereken:
+
+ https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/rmvpe.onnx
+
+```
+
+Intel ARC grafik kartları kullanıcıları Webui'yi başlatmadan önce `source /opt/intel/oneapi/setvars.sh` komutunu çalıştırmalı.
+
+Daha sonra bu komutu kullanarak Webui'yi başlatabilirsiniz:
+```bash
+python web.py
+```
+Windows veya macOS kullanıyorsanız, `RVC-beta.7z` dosyasını indirip çıkararak `go-web.bat`i kullanarak veya macOS'ta `sh ./run.sh` kullanarak doğrudan RVC'yi kullanabilirsiniz.
+
+## Krediler
++ [ContentVec](https://github.com/auspicious3000/contentvec/)
++ [VITS](https://github.com/jaywalnut310/vits)
++ [HIFIGAN](https://github.com/jik876/hifi-gan)
++ [Gradio](https://github.com/gradio-app/gradio)
++ [Ultimate Vocal Remover](https://github.com/Anjok07/ultimatevocalremovergui)
++ [audio-slicer](https://github.com/openvpi/audio-slicer)
++ [Vokal ton çıkarma:RMVPE](https://github.com/Dream-High/RMVPE)
+ + Ön eğitimli model [yxlllc](https://github.com/yxlllc/RMVPE) ve [RVC-Boss](https://github.com/RVC-Boss) tarafından eğitilip test edilmiştir.
+
+## Katkıda Bulunan Herkese Teşekkürler
+[](https://github.com/fumiama/Retrieval-based-Voice-Conversion-WebUI/graphs/contributors)
+```
diff --git a/docs/tr/faiss_tips_tr.md b/docs/tr/faiss_tips_tr.md
new file mode 100644
index 0000000000000000000000000000000000000000..160f36c3f7b1bca296ac9f706db37e6616e90f40
--- /dev/null
+++ b/docs/tr/faiss_tips_tr.md
@@ -0,0 +1,104 @@
+
+# faiss Ayar İpuçları
+==================
+
+# faiss Hakkında
+faiss, yoğun vektörler için komşuluk aramalarının bir kütüphanesidir ve birçok yaklaşık komşuluk arama yöntemini verimli bir şekilde uygular. Facebook araştırma tarafından geliştirilen faiss, benzer vektörleri hızlı bir şekilde bulurken bazı doğruluğu feda eder.
+
+## RVC'de faiss Kullanımı
+RVC'de, HuBERT tarafından dönüştürülen özelliklerin gömülmesi için eğitim verisinden oluşturulan gömme ile benzer gömlemeleri ararız ve bunları karıştırarak orijinal konuşmaya daha yakın bir dönüşüm elde ederiz. Ancak bu arama basitçe yapıldığında zaman alır, bu nedenle yaklaşık komşuluk araması kullanarak yüksek hızlı dönüşüm sağlanır.
+
+# Uygulama Genel Bakış
+Modelin bulunduğu '/logs/your-experiment/3_feature256' dizininde, her ses verisinden HuBERT tarafından çıkarılan özellikler bulunur.
+Buradan, dosya adına göre sıralanmış npy dosyalarını okuyarak vektörleri birleştirip büyük_npy'yi oluştururuz. (Bu vektörün şekli [N, 256] şeklindedir.)
+Büyük_npy'yi /logs/your-experiment/total_fea.npy olarak kaydettikten sonra, onu faiss ile eğitiriz.
+
+Bu makalede, bu parametrelerin anlamını açıklayacağım.
+
+# Yöntemin Açıklaması
+## İndeks Fabrikası
+Bir indeks fabrikası, birden fazla yaklaşık komşuluk arama yöntemini bir dizi olarak bağlayan benzersiz bir faiss gösterimidir. Bu, indeks fabrikası dizesini değiştirerek basitçe çeşitli yaklaşık komşuluk arama yöntemlerini denemenizi sağlar.
+RVC'de bunu şu şekilde kullanırız:
+
+```python
+index = faiss.index_factory(256, "IVF%s,Flat" % n_ivf)
+```
+index_factory'nin argümanları arasında ilk vektör boyutu, ikinci indeks fabrikası dizesi ve üçüncü kullanılacak mesafe yer alır.
+
+Daha ayrıntılı gösterim için
+https://github.com/facebookresearch/faiss/wiki/The-index-factory
+
+## Mesafe İçin İndeks
+Aşağıdaki gibi gömme benzerliği olarak kullanılan iki tipik indeks bulunur.
+
+- Öklidyen mesafe (METRIC_L2)
+- iç çarpım (METRIC_INNER_PRODUCT)
+
+Öklidyen mesafe, her boyutta karesel farkı alır, tüm boyutlardaki farkları toplar ve ardından karekök alır. Bu, günlük hayatta kullandığımız 2D ve 3D'deki mesafeye benzer.
+İç çarpım, çoğunlukla L2 norm ile normalize edildikten sonra iç çarpımı alan ve genellikle kosinüs benzerliği olarak kullanılan bir benzerlik göstergesi olarak kullanılır.
+
+Hangisinin daha iyi olduğu duruma bağlıdır, ancak kosinüs benzerliği genellikle word2vec tarafından elde edilen gömme ve ArcFace tarafından öğrenilen benzer görüntü alım modellerinde kullanılır. Vektör X'i numpy ile l2 normalize yapmak isterseniz, 0 bölme hatasından kaçınmak için yeterince küçük bir eps ile şu kodu kullanabilirsiniz:
+
+```python
+X_normed = X / np.maximum(eps, np.linalg.norm(X, ord=2, axis=-1, keepdims=True))
+```
+
+Ayrıca, indeks fabrikası için üçüncü argüman olarak geçirilecek değeri seçerek hesaplamada kullanılan mesafe indeksini değiştirebilirsiniz.
+
+```python
+index = faiss.index_factory(dimention, text, faiss.METRIC_INNER_PRODUCT)
+```
+
+## IVF
+IVF (Ters dosya indeksleri), tam metin aramasındaki ters indeksle benzer bir algoritmadır.
+Öğrenme sırasında, arama hedefi kmeans ile kümelendirilir ve küme merkezi kullanılarak Voronoi bölütleme gerçekleştirilir. Her veri noktasına bir küme atanır, bu nedenle veri noktalarını kümeden arayan bir sözlük oluştururuz.
+
+Örneğin, kümelere aşağıdaki gibi atanmışsa
+|index|Cluster|
+|-----|-------|
+|1|A|
+|2|B|
+|3|A|
+|4|C|
+|5|B|
+
+Elde edilen ters indeks şu şekildedir:
+
+|cluster|index|
+|-------|-----|
+|A|1, 3|
+|B|2, 5|
+|C|4|
+
+Arama yaparken, önce kümeden n_probe küme ararız ve ardından her küme için ait veri noktalarının mesafelerini hesaplarız.
+
+# Tavsiye Edilen Parametreler
+Resmi olarak nasıl bir indeks seçileceği konusunda rehberler bulunmaktadır, bu nedenle buna uygun olarak açıklayacağım.
+https://github.com/facebookresearch/faiss/wiki/Guidelines-to-choose-an-index
+
+1M'den düşük veri kümeleri için, N sayısı için 4bit-PQ, Nisan 2023 itibariyle faiss'de mevcut en verimli yöntemdir.
+Bunu IVF ile birleştirerek adayları 4bit-PQ ile daraltmak ve nihayet doğru bir indeksle mesafeyi yeniden hesaplamak, aşağıdaki indeks fabrikas
+
+ını kullanarak açıklanabilir.
+
+```python
+index = faiss.index_factory(256, "IVF1024,PQ128x4fs,RFlat")
+```
+
+## IVF İçin Tavsiye Edilen Parametreler
+Çok sayıda IVF durumunu düşünün. Örneğin, veri sayısı için IVF tarafından kabaca nicelleme yapılırsa, bu basit bir tükenmez arama ile aynıdır ve verimsizdir.
+1M veya daha az için IVF değerleri, N veri noktaları için 4*sqrt(N) ~ 16*sqrt(N) arasında tavsiye edilir.
+
+Hesaplama süresi n_probes sayısına orantılı olarak arttığından, doğrulukla danışmanlık yapın ve uygun şekilde seçin. Kişisel olarak, RVC'nin bu kadar doğruluk gerektirmediğini düşünüyorum, bu nedenle n_probe = 1 uygundur.
+
+## FastScan
+FastScan, bunları kaydedicilerde gerçekleştirerek onları Kartez ürünü nicelleme ile hızlı yaklaşık mesafe sağlayan bir yöntemdir.
+Kartez ürünü nicelleme öğrenme sırasında her d boyut için (genellikle d = 2) kümeleme yapar, küme merkezlerini önceden hesaplar ve küme merkezleri arasındaki mesafeyi hesaplar ve bir arama tablosu oluşturur. Tahmin yaparken, her boyutun mesafesi arama tablosuna bakarak O(1) hesaplanabilir.
+PQ sonrası belirttiğiniz sayı genellikle vektörün yarısı olan boyutu belirtir.
+
+FastScan hakkında daha ayrıntılı açıklama için lütfen resmi belgelere başvurun.
+https://github.com/facebookresearch/faiss/wiki/Fast-accumulation-of-PQ-and-AQ-codes-(FastScan)
+
+## RFlat
+RFlat, FastScan ile hesaplanan kesirli mesafeyi indeks fabrikasının üçüncü argümanı tarafından belirtilen doğru mesafe ile yeniden hesaplamak için bir talimattır.
+k komşuları alırken, k*k_factor nokta yeniden hesaplanır.
diff --git a/docs/tr/faq_tr.md b/docs/tr/faq_tr.md
new file mode 100644
index 0000000000000000000000000000000000000000..62b95d0ca9e6f8cabf70faf3a9a6fbc0883628f3
--- /dev/null
+++ b/docs/tr/faq_tr.md
@@ -0,0 +1,98 @@
+## Q1: "Tek Tıklamayla Eğitim" Sonrası İndeks Dosyası Bulunamıyor
+Eğer "Eğitim tamamlandı. Program kapatıldı." mesajını görüyorsa, model başarıyla eğitilmiş demektir ve sonraki hatalar sahte;
+
+"Added" dizini oluşturulduğu halde "Tek Tıklamayla Eğitim" sonrası indeks dosyası bulunamıyorsa, bu genellikle eğitim setinin çok büyük olmasından kaynaklanabilir ve indeksin eklenmesi sıkışabilir. Bu sorun indeks eklerken bellek yükünü azaltmak için toplu işlem yaparak çözülmüştür. Geçici bir çözüm olarak, "Eğitim İndeksini Eğit" düğmesine tekrar tıklamayı deneyin.
+
+## Q2: Eğitim Sonrası "Tonlama İnceleniyor" Bölümünde Model Bulunamıyor
+"Lanetleme İstemi Listesini Yenile" düğmesine tıklayarak tekrar kontrol edin; hala görünmüyorsa, eğitim sırasında herhangi bir hata olup olmadığını kontrol edin ve geliştiricilere daha fazla analiz için konsol, web arayüzü ve logs/experiment_name/*.log ekran görüntülerini gönderin.
+
+## Q3: Bir Model Nasıl Paylaşılır/Başkalarının Modelleri Nasıl Kullanılır?
+rvc_root/logs/experiment_name dizininde saklanan pth dosyaları paylaşım veya çıkarım için değildir, bunlar deney checkpoint'larıdır ve çoğaltılabilirlik ve daha fazla eğitim için saklanır. Paylaşılacak olan model, weights klasöründeki 60+MB'lık pth dosyası olmalıdır;
+
+Gelecekte, weights/exp_name.pth ve logs/exp_name/added_xxx.index birleştirilerek tek bir weights/exp_name.zip dosyasına dönüştürülecek ve manuel indeks girişi gereksinimini ortadan kaldıracaktır; bu nedenle pth dosyasını değil, farklı bir makinede eğitime devam etmek istemezseniz zip dosyasını paylaşın;
+
+Çıkarılmış modelleri zorlama çıkarım için logs klasöründen weights klasörüne birkaç yüz MB'lık pth dosyalarını kopyalamak/paylaşmak, eksik f0, tgt_sr veya diğer anahtarlar gibi hatalara neden olabilir. Smaller modeli manuel veya otomatik olarak çıkarmak için alttaki ckpt sekmesini kullanmanız gerekmektedir (eğer bilgi logs/exp_name içinde bulunuyorsa), pitch bilgisini ve hedef ses örnekleme oranı seçeneklerini seçmeli ve ardından daha küçük modele çıkarmalısınız. Çıkardıktan sonra weights klasöründe 60+ MB'lık bir pth dosyası olacaktır ve sesleri yeniden güncelleyebilirsiniz.
+
+## Q4: Bağlantı Hatası
+Büyük ihtimalle konsolu (siyah komut satırı penceresi) kapatmış olabilirsiniz.
+
+## Q5: Web Arayüzünde 'Beklenen Değer: Satır 1 Sütun 1 (Karakter 0)' Hatası
+Lütfen sistem LAN proxy/global proxy'sini devre dışı bırakın ve ardından sayfayı yenileyin.
+
+## Q6: WebUI Olmadan Nasıl Eğitim Yapılır ve Tahmin Yapılır?
+Eğitim komut dosyası:
+Önce WebUI'de eğitimi çalıştırabilirsiniz, ardından veri seti önişleme ve eğitiminin komut satırı sürümleri mesaj penceresinde görüntülenecektir.
+
+Tahmin komut dosyası:
+https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/myinfer.py
+
+
+örn:
+
+runtime\python.exe myinfer.py 0 "E:\codes\py39\RVC-beta\todo-songs\1111.wav" "E:\codes\py39\logs\mi-test\added_IVF677_Flat_nprobe_7.index" harvest "test.wav" "weights/mi-test.pth" 0.6 cuda:0 True
+
+
+f0up_key=sys.argv[1]
+input_path=sys.argv[2]
+index_path=sys.argv[3]
+f0method=sys.argv[4]#harvest or pm
+opt_path=sys.argv[5]
+model_path=sys.argv[6]
+index_rate=float(sys.argv[7])
+device=sys.argv[8]
+is_half=bool(sys.argv[9])
+
+## Q7: Cuda Hatası/Cuda Bellek Yetersizliği
+Küçük bir ihtimalle CUDA konfigürasyonunda bir problem olabilir veya cihaz desteklenmiyor olabilir; daha muhtemel olarak yetersiz bellek olabilir (bellek yetersizliği).
+
+Eğitim için toplu işlem boyutunu azaltın (1'e indirgemek yeterli değilse, grafik kartını değiştirmeniz gerekebilir); çıkarım için ise config.py dosyasındaki x_pad, x_query, x_center ve x_max ayarlarını ihtiyaca göre düzenleyin. 4GB veya daha düşük bellekli kartlar (örneğin 1060(3G) ve çeşit
+
+li 2GB kartlar) terk edilebilir, 4GB bellekli kartlar hala bir şansı vardır.
+
+## Q8: Optimal Olarak Kaç total_epoch Gerekli?
+Eğitim veri setinin ses kalitesi düşük ve gürültü seviyesi yüksekse, 20-30 dönem yeterlidir. Fazla yüksek bir değer belirlemek, düşük kaliteli eğitim setinizin ses kalitesini artırmaz.
+
+Eğitim setinin ses kalitesi yüksek, gürültü seviyesi düşük ve yeterli süre varsa, bu değeri artırabilirsiniz. 200 kabul edilebilir bir değerdir (çünkü eğitim hızlıdır ve yüksek kaliteli bir eğitim seti hazırlayabiliyorsanız, GPU'nuz muhtemelen uzun bir eğitim süresini sorunsuz bir şekilde yönetebilir).
+
+## Q9: Kaç Dakika Eğitim Verisi Süresi Gerekli?
+
+10 ila 50 dakika arası bir veri seti önerilir.
+
+Garantili yüksek ses kalitesi ve düşük arka plan gürültüsü varsa, veri setinin tonlaması homojen ise daha fazlası eklenebilir.
+
+Yüksek seviyede bir eğitim seti (zarif ve belirgin tonlama), 5 ila 10 dakika arası uygundur.
+
+1 ila 2 dakika veri ile başarılı bir şekilde eğitim yapan bazı insanlar olsa da, başarı diğerleri tarafından tekrarlanabilir değil ve çok bilgilendirici değil. Bu, eğitim setinin çok belirgin bir tonlamaya sahip olmasını (örneğin yüksek frekansta havadar bir anime kız sesi gibi) ve ses kalitesinin yüksek olmasını gerektirir; 1 dakikadan daha kısa süreli veri denenmemiştir ve önerilmez.
+
+
+## Q10: İndeks Oranı Nedir ve Nasıl Ayarlanır?
+Eğer önceden eğitilmiş model ve tahmin kaynağının ton kalitesi, eğitim setinden daha yüksekse, tahmin sonucunun ton kalitesini yükseltebilirler, ancak altta yatan modelin/tahmin kaynağının tonu yerine eğitim setinin tonuna yönelik olası bir ton önyargısıyla sonuçlanır, bu genellikle "ton sızıntısı" olarak adlandırılır.
+
+İndeks oranı, ton sızıntı sorununu azaltmak/çözmek için kullanılır. İndeks oranı 1 olarak ayarlandığında, teorik olarak tahmin kaynağından ton sızıntısı olmaz ve ton kalitesi daha çok eğitim setine yönelik olur. Eğer eğitim seti, tahmin kaynağından daha düşük ses kalitesine sahipse, daha yüksek bir indeks oranı ses kalitesini azaltabilir. Oranı 0'a düşürmek, eğitim seti tonlarını korumak için getirme karıştırmasını kullanmanın etkisine sahip değildir.
+
+Eğer eğitim seti iyi ses kalitesine ve uzun süreye sahipse, total_epoch'u artırın. Model, tahmin kaynağına ve önceden eğitilmiş alt modeline daha az başvurduğunda ve "ton sızıntısı" daha az olduğunda, indeks oranı önemli değil ve hatta indeks dosyası oluşturmak/paylaşmak gerekli değildir.
+
+## Q11: Tahmin Yaparken Hangi GPU'yu Seçmeli?
+config.py dosyasında "device cuda:" ardından kart numarasını seçin.
+
+Kart numarası ile grafik kartı arasındaki eşleme, eğitim sekmesinin grafik kartı bilgileri bölümünde görülebilir.
+
+## Q12: Eğitimin Ortasında Kaydedilen Model Nasıl Kullanılır?
+Kaydetme işlemini ckpt işleme sekmesinin altında yer alan model çıkarımı ile yapabilirsiniz.
+
+## Q13: Dosya/Bellek Hatası (Eğitim Sırasında)?
+Çok fazla işlem ve yetersiz bellek olabilir. Bu sorunu düzeltebilirsiniz:
+
+1. "CPU İş Parçacıkları" alanındaki girişi azaltarak.
+
+2. Eğitim verisini daha kısa ses dosyalarına önceden keserek.
+
+## Q14: Daha Fazla Veri Kullanarak Eğitime Nasıl Devam Edilir?
+
+Adım 1: Tüm wav verilerini path2 dizinine yerleştirin.
+
+Adım 2: exp_name2+path2 -> veri setini önişleme ve özellik çıkarma.
+
+Adım 3: exp_name1 (önceki deneyinizin) en son G ve D dosyalarını exp_name2 klasörüne kopyalayın.
+
+Adım 4: "modeli eğit" düğmesine tıklayın ve önceki deneyinizin model döneminden başlayarak eğitime devam edecektir.
diff --git a/docs/tr/training_tips_tr.md b/docs/tr/training_tips_tr.md
new file mode 100644
index 0000000000000000000000000000000000000000..279fef53ff7c25e1e84e44bb65c0213ba144b0d0
--- /dev/null
+++ b/docs/tr/training_tips_tr.md
@@ -0,0 +1,64 @@
+## RVC Eğitimi için Talimatlar ve İpuçları
+======================================
+Bu TALİMAT, veri eğitiminin nasıl yapıldığını açıklamaktadır.
+
+# Eğitim Akışı
+Eğitim sekmesindeki adımları takip ederek açıklayacağım.
+
+## Adım 1
+Deney adını burada belirleyin.
+
+Ayrıca burada modelin pitch'i dikkate alıp almayacağını da belirleyebilirsiniz.
+Eğer model pitch'i dikkate almazsa, model daha hafif olacak, ancak şarkı söyleme için uygun olmayacaktır.
+
+Her deney için veriler `/logs/your-experiment-name/` dizinine yerleştirilir.
+
+## Adım 2a
+Ses yüklenir ve ön işleme yapılır.
+
+### Ses Yükleme
+Ses içeren bir klasör belirtirseniz, bu klasördeki ses dosyaları otomatik olarak okunur.
+Örneğin, `C:Users\hoge\voices` belirtirseniz, `C:Users\hoge\voices\voice.mp3` yüklenecek, ancak `C:Users\hoge\voices\dir\voice.mp3` yüklenmeyecektir.
+
+### Gürültü Temizleme
+Ses scipy'nin filtfilt işlevi ile yumuşatılır.
+
+### Ses Ayırma
+İlk olarak, giriş sesi belirli bir süreden (max_sil_kept=5 saniye?) daha uzun süren sessiz kısımları tespit ederek böler. Sessizlik üzerinde ses bölündükten sonra sesi 4 saniyede bir 0.3 saniyelik bir örtüşme ile böler. 4 saniye içinde ayrılan sesler için ses normalleştirildikten sonra wav dosyası olarak `/logs/your-experiment-name/0_gt_wavs`'a, ardından 16 kHz örnekleme hızına dönüştürülerek `/logs/your-experiment-name/1_16k_wavs` olarak kaydedilir.
+
+## Adım 2b
+### Pitch Çıkarımı
+Wav dosyalarından pitch bilgisi çıkarılır. ParSelMouth veya PyWorld'e dahili olarak yerleştirilmiş yöntemi kullanarak pitch bilgisi (=f0) çıkarılır ve `/logs/your-experiment-name/2a_f0` dizinine kaydedilir. Ardından pitch bilgisi logaritmik olarak 1 ile 255 arasında bir tamsayıya dönüştürülüp `/logs/your-experiment-name/2b-f0nsf` dizinine kaydedilir.
+
+### Özellik Çıkarımı
+HuBERT'i kullanarak önceden gömme olarak wav dosyasını çıkarır. `/logs/your-experiment-name/1_16k_wavs`'a kaydedilen wav dosyasını okuyarak, wav dosyasını 256 boyutlu HuBERT özelliklerine dönüştürür ve npy formatında `/logs/your-experiment-name/3_feature256` dizinine kaydeder.
+
+## Adım 3
+Modeli eğit.
+### Başlangıç Seviyesi Sözlüğü
+Derin öğrenmede, veri kümesi bölmeye ve öğrenmeye adım adım devam eder. Bir model güncellemesinde (adım), batch_size veri alınır ve tahminler ve hata düzeltmeleri yapılır. Bunun bir defa bir veri kümesi için yapılması bir dönem olarak sayılır.
+
+Bu nedenle, öğrenme zamanı adım başına öğrenme zamanı x (veri kümesindeki veri sayısı / batch boyutu) x dönem sayısıdır. Genel olarak, batch boyutu ne kadar büyükse, öğrenme daha istikrarlı hale gelir (adım başına öğrenme süresi ÷ batch boyutu) küçülür, ancak daha fazla GPU belleği kullanır. GPU RAM'ı nvidia-smi komutu ile kontrol edilebilir. Çalışma ortamının makinesine göre batch boyutunu mümkün olduğunca artırarak öğrenme süresini kısa sürede yapabilirsiniz.
+
+### Önceden Eğitilmiş Modeli Belirtme
+RVC, modeli 0'dan değil önceden eğitilmiş ağırlıklardan başlatarak eğitir, bu nedenle küçük bir veri kümesi ile eğitilebilir.
+
+Varsayılan olarak
+
+- Eğer pitch'i dikkate alıyorsanız, `rvc-location/pretrained/f0G40k.pth` ve `rvc-location/pretrained/f0D40k.pth` yüklenir.
+- Eğer pitch'i dikkate almıyorsanız, yine `rvc-location/pretrained/f0G40k.pth` ve `rvc-location/pretrained/f0D40k.pth` yüklenir.
+
+Öğrenirken model parametreleri her save_every_epoch için `logs/your-experiment-name/G_{}.pth` ve `logs/your-experiment-name/D_{}.pth` olarak kaydedilir, ancak bu yolu belirterek öğrenmeye başlayabilirsiniz. Farklı bir deneyde öğrenilen model ağırlıklarından öğrenmeye yeniden başlayabilir veya eğitimi başlatabilirsiniz.
+
+### Öğrenme İndeksi
+RVC, eğitim sırasında kullanılan HuBERT özellik değerlerini kaydeder ve çıkarım sırasında, öğrenme sırasında kullanılan özellik değerlerine benzer özellik değerlerini arayarak çıkarım yapar. Bu aramayı yüksek hızda gerçekleştirebilmek için indeks öğrenilir.
+İndeks öğrenimi için yaklaş
+
+ık komşuluk arama kütüphanesi faiss kullanılır. `/logs/your-experiment-name/3_feature256`'daki özellik değerini okur ve indeksi öğrenmek için kullanır, `logs/your-experiment-name/add_XXX.index` olarak kaydedilir.
+
+(20230428 güncelleme sürümünden itibaren indeks okunur ve kaydetmek/belirtmek artık gerekli değildir.)
+
+### Düğme Açıklaması
+- Modeli Eğit: Adım 2b'yi çalıştırdıktan sonra, modeli eğitmek için bu düğmeye basın.
+- Özellik İndeksini Eğit: Modeli eğittikten sonra, indeks öğrenme işlemi yapın.
+- Tek Tıklamayla Eğitim: Adım 2b, model eğitimi ve özellik indeks eğitimini bir arada yapar.
diff --git a/gui.py b/gui.py
new file mode 100644
index 0000000000000000000000000000000000000000..0de16b0364f54ecdbca13f1afed3c1a37ed11135
--- /dev/null
+++ b/gui.py
@@ -0,0 +1,1114 @@
+import os
+import sys
+from dotenv import load_dotenv
+import shutil
+
+load_dotenv()
+load_dotenv("sha256.env")
+
+os.environ["OMP_NUM_THREADS"] = "4"
+if sys.platform == "darwin":
+ os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import multiprocessing
+
+flag_vc = False
+
+
+def printt(strr, *args):
+ if len(args) == 0:
+ print(strr)
+ else:
+ print(strr % args)
+
+
+def phase_vocoder(a, b, fade_out, fade_in):
+ window = torch.sqrt(fade_out * fade_in)
+ fa = torch.fft.rfft(a * window)
+ fb = torch.fft.rfft(b * window)
+ absab = torch.abs(fa) + torch.abs(fb)
+ n = a.shape[0]
+ if n % 2 == 0:
+ absab[1:-1] *= 2
+ else:
+ absab[1:] *= 2
+ phia = torch.angle(fa)
+ phib = torch.angle(fb)
+ deltaphase = phib - phia
+ deltaphase = deltaphase - 2 * np.pi * torch.floor(deltaphase / 2 / np.pi + 0.5)
+ w = 2 * np.pi * torch.arange(n // 2 + 1).to(a) + deltaphase
+ t = torch.arange(n).unsqueeze(-1).to(a) / n
+ result = (
+ a * (fade_out**2)
+ + b * (fade_in**2)
+ + torch.sum(absab * torch.cos(w * t + phia), -1) * window / n
+ )
+ return result
+
+
+class Harvest(multiprocessing.Process):
+ def __init__(self, inp_q, opt_q):
+ multiprocessing.Process.__init__(self)
+ self.inp_q = inp_q
+ self.opt_q = opt_q
+
+ def run(self):
+ import numpy as np
+ import pyworld
+
+ while 1:
+ idx, x, res_f0, n_cpu, ts = self.inp_q.get()
+ f0, t = pyworld.harvest(
+ x.astype(np.double),
+ fs=16000,
+ f0_ceil=1100,
+ f0_floor=50,
+ frame_period=10,
+ )
+ res_f0[idx] = f0
+ if len(res_f0.keys()) >= n_cpu:
+ self.opt_q.put(ts)
+
+
+if __name__ == "__main__":
+ import json
+ import multiprocessing
+ import re
+ import time
+ from multiprocessing import Queue, cpu_count
+
+ import librosa
+ from infer.modules.gui import TorchGate
+ import numpy as np
+ import FreeSimpleGUI as sg
+ import sounddevice as sd
+ import torch
+ import torch.nn.functional as F
+ import torchaudio.transforms as tat
+
+ import infer.lib.rtrvc as rtrvc
+ from i18n.i18n import I18nAuto
+ from configs import Config
+
+ i18n = I18nAuto()
+
+ # device = rvc_for_realtime.config.device
+ # device = torch.device(
+ # "cuda"
+ # if torch.cuda.is_available()
+ # else ("mps" if torch.backends.mps.is_available() else "cpu")
+ # )
+ current_dir = os.getcwd()
+ inp_q = Queue()
+ opt_q = Queue()
+ n_cpu = min(cpu_count(), 8)
+ for _ in range(n_cpu):
+ p = Harvest(inp_q, opt_q)
+ p.daemon = True
+ p.start()
+
+ class GUIConfig:
+ def __init__(self) -> None:
+ self.pth_path: str = ""
+ self.index_path: str = ""
+ self.pitch: int = 0
+ self.formant: float = 0.0
+ self.sr_type: str = "sr_model"
+ self.block_time: float = 0.25 # s
+ self.threhold: int = -60
+ self.crossfade_time: float = 0.05
+ self.extra_time: float = 2.5
+ self.I_noise_reduce: bool = False
+ self.O_noise_reduce: bool = False
+ self.use_pv: bool = False
+ self.rms_mix_rate: float = 0.0
+ self.index_rate: float = 0.0
+ self.n_cpu: int = min(n_cpu, 4)
+ self.f0method: str = "fcpe"
+ self.sg_hostapi: str = ""
+ self.wasapi_exclusive: bool = False
+ self.sg_input_device: str = ""
+ self.sg_output_device: str = ""
+
+ class GUI:
+ def __init__(self) -> None:
+ self.gui_config = GUIConfig()
+ self.config = Config()
+ self.function = "vc"
+ self.delay_time = 0
+ self.hostapis = None
+ self.input_devices = None
+ self.output_devices = None
+ self.input_devices_indices = None
+ self.output_devices_indices = None
+ self.stream = None
+ self.update_devices()
+ self.launcher()
+
+ def check_assets(self):
+ global now_dir
+ from infer.lib.rvcmd import check_all_assets, download_all_assets
+
+ tmp = os.path.join(now_dir, "TEMP")
+ shutil.rmtree(tmp, ignore_errors=True)
+ os.makedirs(tmp, exist_ok=True)
+ if not check_all_assets(update=self.config.update):
+ if self.config.update:
+ download_all_assets(tmpdir=tmp)
+ if not check_all_assets(update=self.config.update):
+ printt("counld not satisfy all assets needed.")
+ exit(1)
+
+ def load(self):
+ try:
+ if not os.path.exists("configs/inuse/config.json"):
+ shutil.copy("configs/config.json", "configs/inuse/config.json")
+ with open("configs/inuse/config.json", "r") as j:
+ data = json.load(j)
+ data["sr_model"] = data["sr_type"] == "sr_model"
+ data["sr_device"] = data["sr_type"] == "sr_device"
+ data["pm"] = data["f0method"] == "pm"
+ data["dio"] = data["f0method"] == "dio"
+ data["harvest"] = data["f0method"] == "harvest"
+ data["crepe"] = data["f0method"] == "crepe"
+ data["rmvpe"] = data["f0method"] == "rmvpe"
+ data["fcpe"] = data["f0method"] == "fcpe"
+ if data["sg_hostapi"] in self.hostapis:
+ self.update_devices(hostapi_name=data["sg_hostapi"])
+ if (
+ data["sg_input_device"] not in self.input_devices
+ or data["sg_output_device"] not in self.output_devices
+ ):
+ self.update_devices()
+ data["sg_hostapi"] = self.hostapis[0]
+ data["sg_input_device"] = self.input_devices[
+ self.input_devices_indices.index(sd.default.device[0])
+ ]
+ data["sg_output_device"] = self.output_devices[
+ self.output_devices_indices.index(sd.default.device[1])
+ ]
+ else:
+ data["sg_hostapi"] = self.hostapis[0]
+ data["sg_input_device"] = self.input_devices[
+ self.input_devices_indices.index(sd.default.device[0])
+ ]
+ data["sg_output_device"] = self.output_devices[
+ self.output_devices_indices.index(sd.default.device[1])
+ ]
+ except:
+ with open("configs/inuse/config.json", "w") as j:
+ data = {
+ "pth_path": "",
+ "index_path": "",
+ "sg_hostapi": self.hostapis[0],
+ "sg_wasapi_exclusive": False,
+ "sg_input_device": self.input_devices[
+ self.input_devices_indices.index(sd.default.device[0])
+ ],
+ "sg_output_device": self.output_devices[
+ self.output_devices_indices.index(sd.default.device[1])
+ ],
+ "sr_type": "sr_model",
+ "threhold": -60,
+ "pitch": 0,
+ "formant": 0.0,
+ "index_rate": 0,
+ "rms_mix_rate": 0,
+ "block_time": 0.25,
+ "crossfade_length": 0.05,
+ "extra_time": 2.5,
+ "n_cpu": 4,
+ "f0method": "rmvpe",
+ "use_jit": False,
+ "use_pv": False,
+ }
+ data["sr_model"] = data["sr_type"] == "sr_model"
+ data["sr_device"] = data["sr_type"] == "sr_device"
+ data["pm"] = data["f0method"] == "pm"
+ data["dio"] = data["f0method"] == "dio"
+ data["harvest"] = data["f0method"] == "harvest"
+ data["crepe"] = data["f0method"] == "crepe"
+ data["rmvpe"] = data["f0method"] == "rmvpe"
+ data["fcpe"] = data["f0method"] == "fcpe"
+ return data
+
+ def launcher(self):
+ data = self.load()
+ self.config.use_jit = False # data.get("use_jit", self.config.use_jit)
+ sg.theme("LightBlue3")
+ layout = [
+ [
+ sg.Frame(
+ title=i18n("Load model"),
+ layout=[
+ [
+ sg.Input(
+ default_text=data.get("pth_path", ""),
+ key="pth_path",
+ ),
+ sg.FileBrowse(
+ i18n("Select the .pth file"),
+ initial_folder=os.path.join(
+ os.getcwd(), "assets/weights"
+ ),
+ file_types=[("Model File", "*.pth")],
+ ),
+ ],
+ [
+ sg.Input(
+ default_text=data.get("index_path", ""),
+ key="index_path",
+ ),
+ sg.FileBrowse(
+ i18n("Select the .index file"),
+ initial_folder=os.path.join(os.getcwd(), "logs"),
+ file_types=[("Index File", "*.index")],
+ ),
+ ],
+ ],
+ )
+ ],
+ [
+ sg.Frame(
+ layout=[
+ [
+ sg.Text(i18n("Device type")),
+ sg.Combo(
+ self.hostapis,
+ key="sg_hostapi",
+ default_value=data.get("sg_hostapi", ""),
+ enable_events=True,
+ size=(20, 1),
+ ),
+ sg.Checkbox(
+ i18n("Takeover WASAPI device"),
+ key="sg_wasapi_exclusive",
+ default=data.get("sg_wasapi_exclusive", False),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Input device")),
+ sg.Combo(
+ self.input_devices,
+ key="sg_input_device",
+ default_value=data.get("sg_input_device", ""),
+ enable_events=True,
+ size=(45, 1),
+ ),
+ ],
+ [
+ sg.Text(i18n("Output device")),
+ sg.Combo(
+ self.output_devices,
+ key="sg_output_device",
+ default_value=data.get("sg_output_device", ""),
+ enable_events=True,
+ size=(45, 1),
+ ),
+ ],
+ [
+ sg.Button(
+ i18n("Reload device list"), key="reload_devices"
+ ),
+ sg.Radio(
+ i18n("Choose sample rate of the model"),
+ "sr_type",
+ key="sr_model",
+ default=data.get("sr_model", True),
+ enable_events=True,
+ ),
+ sg.Radio(
+ i18n("Choose sample rate of the device"),
+ "sr_type",
+ key="sr_device",
+ default=data.get("sr_device", False),
+ enable_events=True,
+ ),
+ sg.Text(i18n("Sampling rate")),
+ sg.Text("", key="sr_stream"),
+ ],
+ ],
+ title=i18n("Audio device"),
+ )
+ ],
+ [
+ sg.Frame(
+ layout=[
+ [
+ sg.Text(i18n("Response threshold")),
+ sg.Slider(
+ range=(-60, 0),
+ key="threhold",
+ resolution=1,
+ orientation="h",
+ default_value=data.get("threhold", -60),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Pitch settings")),
+ sg.Slider(
+ range=(-24, 24),
+ key="pitch",
+ resolution=1,
+ orientation="h",
+ default_value=data.get("pitch", 0),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Formant offset")),
+ sg.Slider(
+ range=(-5, 5),
+ key="formant",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("formant", 0.0),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Feature searching ratio")),
+ sg.Slider(
+ range=(0.0, 1.0),
+ key="index_rate",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("index_rate", 0),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Loudness factor")),
+ sg.Slider(
+ range=(0.0, 1.0),
+ key="rms_mix_rate",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("rms_mix_rate", 0),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Pitch detection algorithm")),
+ sg.Radio(
+ "pm",
+ "f0method",
+ key="pm",
+ default=data.get("pm", False),
+ enable_events=True,
+ ),
+ sg.Radio(
+ "dio",
+ "f0method",
+ key="dio",
+ default=data.get("dio", False),
+ enable_events=True,
+ ),
+ sg.Radio(
+ "harvest",
+ "f0method",
+ key="harvest",
+ default=data.get("harvest", False),
+ enable_events=True,
+ ),
+ sg.Radio(
+ "crepe",
+ "f0method",
+ key="crepe",
+ default=data.get("crepe", False),
+ enable_events=True,
+ ),
+ sg.Radio(
+ "rmvpe",
+ "f0method",
+ key="rmvpe",
+ default=data.get("rmvpe", False),
+ enable_events=True,
+ ),
+ sg.Radio(
+ "fcpe",
+ "f0method",
+ key="fcpe",
+ default=data.get("fcpe", True),
+ enable_events=True,
+ ),
+ ],
+ ],
+ title=i18n("General settings"),
+ ),
+ sg.Frame(
+ layout=[
+ [
+ sg.Text(i18n("Sample length")),
+ sg.Slider(
+ range=(0.02, 1.5),
+ key="block_time",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("block_time", 0.25),
+ enable_events=True,
+ ),
+ ],
+ # [
+ # sg.Text("设备延迟"),
+ # sg.Slider(
+ # range=(0, 1),
+ # key="device_latency",
+ # resolution=0.001,
+ # orientation="h",
+ # default_value=data.get("device_latency", 0.1),
+ # enable_events=True,
+ # ),
+ # ],
+ [
+ sg.Text(
+ i18n(
+ "Number of CPU processes used for harvest pitch algorithm"
+ )
+ ),
+ sg.Slider(
+ range=(1, n_cpu),
+ key="n_cpu",
+ resolution=1,
+ orientation="h",
+ default_value=data.get(
+ "n_cpu", min(self.gui_config.n_cpu, n_cpu)
+ ),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Fade length")),
+ sg.Slider(
+ range=(0.01, 0.15),
+ key="crossfade_length",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("crossfade_length", 0.05),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Text(i18n("Extra inference time")),
+ sg.Slider(
+ range=(0.05, 5.00),
+ key="extra_time",
+ resolution=0.01,
+ orientation="h",
+ default_value=data.get("extra_time", 2.5),
+ enable_events=True,
+ ),
+ ],
+ [
+ sg.Checkbox(
+ i18n("Input noise reduction"),
+ key="I_noise_reduce",
+ enable_events=True,
+ ),
+ sg.Checkbox(
+ i18n("Output noise reduction"),
+ key="O_noise_reduce",
+ enable_events=True,
+ ),
+ sg.Checkbox(
+ i18n("Enable phase vocoder"),
+ key="use_pv",
+ default=data.get("use_pv", False),
+ enable_events=True,
+ ),
+ # sg.Checkbox(
+ # "JIT加速",
+ # default=self.config.use_jit,
+ # key="use_jit",
+ # enable_events=False,
+ # ),
+ ],
+ # [sg.Text("注：首次使用JIT加速时，会出现卡顿，\n 并伴随一些噪音，但这是正常现象！")],
+ ],
+ title=i18n("Performance settings"),
+ ),
+ ],
+ [
+ sg.Button(i18n("Start audio conversion"), key="start_vc"),
+ sg.Button(i18n("Stop audio conversion"), key="stop_vc"),
+ sg.Radio(
+ i18n("Input voice monitor"),
+ "function",
+ key="im",
+ default=False,
+ enable_events=True,
+ ),
+ sg.Radio(
+ i18n("Output converted voice"),
+ "function",
+ key="vc",
+ default=True,
+ enable_events=True,
+ ),
+ sg.Text(i18n("Algorithmic delays (ms)")),
+ sg.Text("0", key="delay_time"),
+ sg.Text(i18n("Inference time (ms)")),
+ sg.Text("0", key="infer_time"),
+ ],
+ ]
+ self.window = sg.Window("RVC - GUI", layout=layout, finalize=True)
+ self.event_handler()
+
+ def event_handler(self):
+ global flag_vc
+ while True:
+ event, values = self.window.read()
+ if event == sg.WINDOW_CLOSED:
+ self.stop_stream()
+ exit()
+ if event == "reload_devices" or event == "sg_hostapi":
+ self.gui_config.sg_hostapi = values["sg_hostapi"]
+ self.update_devices(hostapi_name=values["sg_hostapi"])
+ if self.gui_config.sg_hostapi not in self.hostapis:
+ self.gui_config.sg_hostapi = self.hostapis[0]
+ self.window["sg_hostapi"].Update(values=self.hostapis)
+ self.window["sg_hostapi"].Update(value=self.gui_config.sg_hostapi)
+ if (
+ self.gui_config.sg_input_device not in self.input_devices
+ and len(self.input_devices) > 0
+ ):
+ self.gui_config.sg_input_device = self.input_devices[0]
+ self.window["sg_input_device"].Update(values=self.input_devices)
+ self.window["sg_input_device"].Update(
+ value=self.gui_config.sg_input_device
+ )
+ if self.gui_config.sg_output_device not in self.output_devices:
+ self.gui_config.sg_output_device = self.output_devices[0]
+ self.window["sg_output_device"].Update(values=self.output_devices)
+ self.window["sg_output_device"].Update(
+ value=self.gui_config.sg_output_device
+ )
+ if event == "start_vc" and not flag_vc:
+ if self.set_values(values) == True:
+ printt("cuda_is_available: %s", torch.cuda.is_available())
+ self.start_vc()
+ settings = {
+ "pth_path": values["pth_path"],
+ "index_path": values["index_path"],
+ "sg_hostapi": values["sg_hostapi"],
+ "sg_wasapi_exclusive": values["sg_wasapi_exclusive"],
+ "sg_input_device": values["sg_input_device"],
+ "sg_output_device": values["sg_output_device"],
+ "sr_type": ["sr_model", "sr_device"][
+ [
+ values["sr_model"],
+ values["sr_device"],
+ ].index(True)
+ ],
+ "threhold": values["threhold"],
+ "pitch": values["pitch"],
+ "formant": values["formant"],
+ "rms_mix_rate": values["rms_mix_rate"],
+ "index_rate": values["index_rate"],
+ # "device_latency": values["device_latency"],
+ "block_time": values["block_time"],
+ "crossfade_length": values["crossfade_length"],
+ "extra_time": values["extra_time"],
+ "n_cpu": values["n_cpu"],
+ # "use_jit": values["use_jit"],
+ "use_jit": False,
+ "use_pv": values["use_pv"],
+ "f0method": [
+ "pm",
+ "dio",
+ "harvest",
+ "crepe",
+ "rmvpe",
+ "fcpe",
+ ][
+ [
+ values["pm"],
+ values["dio"],
+ values["harvest"],
+ values["crepe"],
+ values["rmvpe"],
+ values["fcpe"],
+ ].index(True)
+ ],
+ }
+ with open("configs/inuse/config.json", "w") as j:
+ json.dump(settings, j)
+ if self.stream is not None:
+ self.delay_time = (
+ self.stream.latency[-1]
+ + values["block_time"]
+ + values["crossfade_length"]
+ + 0.01
+ )
+ if values["I_noise_reduce"]:
+ self.delay_time += min(values["crossfade_length"], 0.04)
+ self.window["sr_stream"].update(self.gui_config.samplerate)
+ self.window["delay_time"].update(
+ int(np.round(self.delay_time * 1000))
+ )
+ # Parameter hot update
+ if event == "threhold":
+ self.gui_config.threhold = values["threhold"]
+ elif event == "pitch":
+ self.gui_config.pitch = values["pitch"]
+ if hasattr(self, "rvc"):
+ self.rvc.set_key(values["pitch"])
+ elif event == "formant":
+ self.gui_config.formant = values["formant"]
+ if hasattr(self, "rvc"):
+ self.rvc.set_formant(values["formant"])
+ elif event == "index_rate":
+ self.gui_config.index_rate = values["index_rate"]
+ if hasattr(self, "rvc"):
+ self.rvc.set_index_rate(values["index_rate"])
+ elif event == "rms_mix_rate":
+ self.gui_config.rms_mix_rate = values["rms_mix_rate"]
+ elif event in ["pm", "dio", "harvest", "crepe", "rmvpe", "fcpe"]:
+ self.gui_config.f0method = event
+ elif event == "I_noise_reduce":
+ self.gui_config.I_noise_reduce = values["I_noise_reduce"]
+ if self.stream is not None:
+ self.delay_time += (
+ 1 if values["I_noise_reduce"] else -1
+ ) * min(values["crossfade_length"], 0.04)
+ self.window["delay_time"].update(
+ int(np.round(self.delay_time * 1000))
+ )
+ elif event == "O_noise_reduce":
+ self.gui_config.O_noise_reduce = values["O_noise_reduce"]
+ elif event == "use_pv":
+ self.gui_config.use_pv = values["use_pv"]
+ elif event in ["vc", "im"]:
+ self.function = event
+ elif event == "stop_vc" or event != "start_vc":
+ # Other parameters do not support hot update
+ self.stop_stream()
+
+ def set_values(self, values):
+ if len(values["pth_path"].strip()) == 0:
+ sg.popup(i18n("Please choose the .pth file"))
+ return False
+ if len(values["index_path"].strip()) == 0:
+ sg.popup(i18n("Please choose the .index file"))
+ return False
+ pattern = re.compile("[^\x00-\x7F]+")
+ if pattern.findall(values["pth_path"]):
+ sg.popup(i18n("pth path cannot contain unicode characters"))
+ return False
+ if pattern.findall(values["index_path"]):
+ sg.popup(i18n("index path cannot contain unicode characters"))
+ return False
+ self.set_devices(values["sg_input_device"], values["sg_output_device"])
+ self.config.use_jit = False # values["use_jit"]
+ # self.device_latency = values["device_latency"]
+ self.gui_config.sg_hostapi = values["sg_hostapi"]
+ self.gui_config.sg_wasapi_exclusive = values["sg_wasapi_exclusive"]
+ self.gui_config.sg_input_device = values["sg_input_device"]
+ self.gui_config.sg_output_device = values["sg_output_device"]
+ self.gui_config.pth_path = values["pth_path"]
+ self.gui_config.index_path = values["index_path"]
+ self.gui_config.sr_type = ["sr_model", "sr_device"][
+ [
+ values["sr_model"],
+ values["sr_device"],
+ ].index(True)
+ ]
+ self.gui_config.threhold = values["threhold"]
+ self.gui_config.pitch = values["pitch"]
+ self.gui_config.formant = values["formant"]
+ self.gui_config.block_time = values["block_time"]
+ self.gui_config.crossfade_time = values["crossfade_length"]
+ self.gui_config.extra_time = values["extra_time"]
+ self.gui_config.I_noise_reduce = values["I_noise_reduce"]
+ self.gui_config.O_noise_reduce = values["O_noise_reduce"]
+ self.gui_config.use_pv = values["use_pv"]
+ self.gui_config.rms_mix_rate = values["rms_mix_rate"]
+ self.gui_config.index_rate = values["index_rate"]
+ self.gui_config.n_cpu = values["n_cpu"]
+ self.gui_config.f0method = [
+ "pm",
+ "dio",
+ "harvest",
+ "crepe",
+ "rmvpe",
+ "fcpe",
+ ][
+ [
+ values["pm"],
+ values["dio"],
+ values["harvest"],
+ values["crepe"],
+ values["rmvpe"],
+ values["fcpe"],
+ ].index(True)
+ ]
+ return True
+
+ def start_vc(self):
+ torch.cuda.empty_cache()
+ self.rvc = rtrvc.RVC(
+ self.gui_config.pitch,
+ self.gui_config.formant,
+ self.gui_config.pth_path,
+ self.gui_config.index_path,
+ self.gui_config.index_rate,
+ self.gui_config.n_cpu,
+ self.config.device,
+ self.config.use_jit,
+ self.config.is_half,
+ self.config.dml,
+ )
+ self.gui_config.samplerate = (
+ self.rvc.tgt_sr
+ if self.gui_config.sr_type == "sr_model"
+ else self.get_device_samplerate()
+ )
+ self.gui_config.channels = self.get_device_channels()
+ self.zc = self.gui_config.samplerate // 100
+ self.block_frame = (
+ int(
+ np.round(
+ self.gui_config.block_time
+ * self.gui_config.samplerate
+ / self.zc
+ )
+ )
+ * self.zc
+ )
+ self.block_frame_16k = 160 * self.block_frame // self.zc
+ self.crossfade_frame = (
+ int(
+ np.round(
+ self.gui_config.crossfade_time
+ * self.gui_config.samplerate
+ / self.zc
+ )
+ )
+ * self.zc
+ )
+ self.sola_buffer_frame = min(self.crossfade_frame, 4 * self.zc)
+ self.sola_search_frame = self.zc
+ self.extra_frame = (
+ int(
+ np.round(
+ self.gui_config.extra_time
+ * self.gui_config.samplerate
+ / self.zc
+ )
+ )
+ * self.zc
+ )
+ self.input_wav: torch.Tensor = torch.zeros(
+ self.extra_frame
+ + self.crossfade_frame
+ + self.sola_search_frame
+ + self.block_frame,
+ device=self.config.device,
+ dtype=torch.float32,
+ )
+ self.input_wav_denoise: torch.Tensor = self.input_wav.clone()
+ self.input_wav_res: torch.Tensor = torch.zeros(
+ 160 * self.input_wav.shape[0] // self.zc,
+ device=self.config.device,
+ dtype=torch.float32,
+ )
+ self.rms_buffer: np.ndarray = np.zeros(4 * self.zc, dtype="float32")
+ self.sola_buffer: torch.Tensor = torch.zeros(
+ self.sola_buffer_frame, device=self.config.device, dtype=torch.float32
+ )
+ self.nr_buffer: torch.Tensor = self.sola_buffer.clone()
+ self.output_buffer: torch.Tensor = self.input_wav.clone()
+ self.skip_head = self.extra_frame // self.zc
+ self.return_length = (
+ self.block_frame + self.sola_buffer_frame + self.sola_search_frame
+ ) // self.zc
+ self.fade_in_window: torch.Tensor = (
+ torch.sin(
+ 0.5
+ * np.pi
+ * torch.linspace(
+ 0.0,
+ 1.0,
+ steps=self.sola_buffer_frame,
+ device=self.config.device,
+ dtype=torch.float32,
+ )
+ )
+ ** 2
+ )
+ self.fade_out_window: torch.Tensor = 1 - self.fade_in_window
+ self.resampler = tat.Resample(
+ orig_freq=self.gui_config.samplerate,
+ new_freq=16000,
+ dtype=torch.float32,
+ ).to(self.config.device)
+ if self.rvc.tgt_sr != self.gui_config.samplerate:
+ self.resampler2 = tat.Resample(
+ orig_freq=self.rvc.tgt_sr,
+ new_freq=self.gui_config.samplerate,
+ dtype=torch.float32,
+ ).to(self.config.device)
+ else:
+ self.resampler2 = None
+ self.tg = TorchGate(
+ sr=self.gui_config.samplerate, n_fft=4 * self.zc, prop_decrease=0.9
+ ).to(self.config.device)
+ self.start_stream()
+
+ def start_stream(self):
+ global flag_vc
+ if not flag_vc:
+ flag_vc = True
+ if (
+ "WASAPI" in self.gui_config.sg_hostapi
+ and self.gui_config.sg_wasapi_exclusive
+ ):
+ extra_settings = sd.WasapiSettings(exclusive=True)
+ else:
+ extra_settings = None
+ self.stream = sd.Stream(
+ callback=self.audio_callback,
+ blocksize=self.block_frame,
+ samplerate=self.gui_config.samplerate,
+ channels=self.gui_config.channels,
+ dtype="float32",
+ extra_settings=extra_settings,
+ )
+ self.stream.start()
+
+ def stop_stream(self):
+ global flag_vc
+ if flag_vc:
+ flag_vc = False
+ if self.stream is not None:
+ self.stream.abort()
+ self.stream.close()
+ self.stream = None
+
+ def audio_callback(
+ self, indata: np.ndarray, outdata: np.ndarray, frames, times, status
+ ):
+ """
+ 音频处理
+ """
+ global flag_vc
+ start_time = time.perf_counter()
+ indata = librosa.to_mono(indata.T)
+ if self.gui_config.threhold > -60:
+ indata = np.append(self.rms_buffer, indata)
+ rms = librosa.feature.rms(
+ y=indata, frame_length=4 * self.zc, hop_length=self.zc
+ )[:, 2:]
+ self.rms_buffer[:] = indata[-4 * self.zc :]
+ indata = indata[2 * self.zc - self.zc // 2 :]
+ db_threhold = (
+ librosa.amplitude_to_db(rms, ref=1.0)[0] < self.gui_config.threhold
+ )
+ for i in range(db_threhold.shape[0]):
+ if db_threhold[i]:
+ indata[i * self.zc : (i + 1) * self.zc] = 0
+ indata = indata[self.zc // 2 :]
+ self.input_wav[: -self.block_frame] = self.input_wav[
+ self.block_frame :
+ ].clone()
+ self.input_wav[-indata.shape[0] :] = torch.from_numpy(indata).to(
+ self.config.device
+ )
+ self.input_wav_res[: -self.block_frame_16k] = self.input_wav_res[
+ self.block_frame_16k :
+ ].clone()
+ # input noise reduction and resampling
+ if self.gui_config.I_noise_reduce:
+ self.input_wav_denoise[: -self.block_frame] = self.input_wav_denoise[
+ self.block_frame :
+ ].clone()
+ input_wav = self.input_wav[-self.sola_buffer_frame - self.block_frame :]
+ input_wav = self.tg(
+ input_wav.unsqueeze(0), self.input_wav.unsqueeze(0)
+ ).squeeze(0)
+ input_wav[: self.sola_buffer_frame] *= self.fade_in_window
+ input_wav[: self.sola_buffer_frame] += (
+ self.nr_buffer * self.fade_out_window
+ )
+ self.input_wav_denoise[-self.block_frame :] = input_wav[
+ : self.block_frame
+ ]
+ self.nr_buffer[:] = input_wav[self.block_frame :]
+ self.input_wav_res[-self.block_frame_16k - 160 :] = self.resampler(
+ self.input_wav_denoise[-self.block_frame - 2 * self.zc :]
+ )[160:]
+ else:
+ self.input_wav_res[-160 * (indata.shape[0] // self.zc + 1) :] = (
+ self.resampler(self.input_wav[-indata.shape[0] - 2 * self.zc :])[
+ 160:
+ ]
+ )
+ # infer
+ if self.function == "vc":
+ infer_wav = self.rvc.infer(
+ self.input_wav_res,
+ self.block_frame_16k,
+ self.skip_head,
+ self.return_length,
+ self.gui_config.f0method,
+ )
+ if self.resampler2 is not None:
+ infer_wav = self.resampler2(infer_wav)
+ elif self.gui_config.I_noise_reduce:
+ infer_wav = self.input_wav_denoise[self.extra_frame :].clone()
+ else:
+ infer_wav = self.input_wav[self.extra_frame :].clone()
+ # output noise reduction
+ if self.gui_config.O_noise_reduce and self.function == "vc":
+ self.output_buffer[: -self.block_frame] = self.output_buffer[
+ self.block_frame :
+ ].clone()
+ self.output_buffer[-self.block_frame :] = infer_wav[-self.block_frame :]
+ infer_wav = self.tg(
+ infer_wav.unsqueeze(0), self.output_buffer.unsqueeze(0)
+ ).squeeze(0)
+ # volume envelop mixing
+ if self.gui_config.rms_mix_rate < 1 and self.function == "vc":
+ if self.gui_config.I_noise_reduce:
+ input_wav = self.input_wav_denoise[self.extra_frame :]
+ else:
+ input_wav = self.input_wav[self.extra_frame :]
+ rms1 = librosa.feature.rms(
+ y=input_wav[: infer_wav.shape[0]].cpu().numpy(),
+ frame_length=4 * self.zc,
+ hop_length=self.zc,
+ )
+ rms1 = torch.from_numpy(rms1).to(self.config.device)
+ rms1 = F.interpolate(
+ rms1.unsqueeze(0),
+ size=infer_wav.shape[0] + 1,
+ mode="linear",
+ align_corners=True,
+ )[0, 0, :-1]
+ rms2 = librosa.feature.rms(
+ y=infer_wav[:].cpu().numpy(),
+ frame_length=4 * self.zc,
+ hop_length=self.zc,
+ )
+ rms2 = torch.from_numpy(rms2).to(self.config.device)
+ rms2 = F.interpolate(
+ rms2.unsqueeze(0),
+ size=infer_wav.shape[0] + 1,
+ mode="linear",
+ align_corners=True,
+ )[0, 0, :-1]
+ rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-3)
+ infer_wav *= torch.pow(
+ rms1 / rms2, torch.tensor(1 - self.gui_config.rms_mix_rate)
+ )
+ # SOLA algorithm from https://github.com/yxlllc/DDSP-SVC
+ conv_input = infer_wav[
+ None, None, : self.sola_buffer_frame + self.sola_search_frame
+ ]
+ cor_nom = F.conv1d(conv_input, self.sola_buffer[None, None, :])
+ cor_den = torch.sqrt(
+ F.conv1d(
+ conv_input**2,
+ torch.ones(1, 1, self.sola_buffer_frame, device=self.config.device),
+ )
+ + 1e-8
+ )
+ if sys.platform == "darwin":
+ _, sola_offset = torch.max(cor_nom[0, 0] / cor_den[0, 0])
+ sola_offset = sola_offset.item()
+ else:
+ sola_offset = torch.argmax(cor_nom[0, 0] / cor_den[0, 0])
+ # printt("sola_offset = %d", int(sola_offset))
+ infer_wav = infer_wav[sola_offset:]
+ if "privateuseone" in str(self.config.device) or not self.gui_config.use_pv:
+ infer_wav[: self.sola_buffer_frame] *= self.fade_in_window
+ infer_wav[: self.sola_buffer_frame] += (
+ self.sola_buffer * self.fade_out_window
+ )
+ else:
+ infer_wav[: self.sola_buffer_frame] = phase_vocoder(
+ self.sola_buffer,
+ infer_wav[: self.sola_buffer_frame],
+ self.fade_out_window,
+ self.fade_in_window,
+ )
+ self.sola_buffer[:] = infer_wav[
+ self.block_frame : self.block_frame + self.sola_buffer_frame
+ ]
+ outdata[:] = (
+ infer_wav[: self.block_frame]
+ .repeat(self.gui_config.channels, 1)
+ .t()
+ .cpu()
+ .numpy()
+ )
+ total_time = time.perf_counter() - start_time
+ if flag_vc:
+ self.window["infer_time"].update(int(total_time * 1000))
+ # printt("Infer time: %.2f", total_time)
+
+ def update_devices(self, hostapi_name=None):
+ """获取设备列表"""
+ global flag_vc
+ flag_vc = False
+ sd._terminate()
+ sd._initialize()
+ devices = sd.query_devices()
+ hostapis = sd.query_hostapis()
+ for hostapi in hostapis:
+ for device_idx in hostapi["devices"]:
+ devices[device_idx]["hostapi_name"] = hostapi["name"]
+ self.hostapis = [hostapi["name"] for hostapi in hostapis]
+ if hostapi_name not in self.hostapis:
+ hostapi_name = self.hostapis[0]
+ self.input_devices = [
+ d["name"]
+ for d in devices
+ if d["max_input_channels"] > 0 and d["hostapi_name"] == hostapi_name
+ ]
+ self.output_devices = [
+ d["name"]
+ for d in devices
+ if d["max_output_channels"] > 0 and d["hostapi_name"] == hostapi_name
+ ]
+ self.input_devices_indices = [
+ d["index"] if "index" in d else d["name"]
+ for d in devices
+ if d["max_input_channels"] > 0 and d["hostapi_name"] == hostapi_name
+ ]
+ self.output_devices_indices = [
+ d["index"] if "index" in d else d["name"]
+ for d in devices
+ if d["max_output_channels"] > 0 and d["hostapi_name"] == hostapi_name
+ ]
+
+ def set_devices(self, input_device, output_device):
+ """设置输出设备"""
+ sd.default.device[0] = self.input_devices_indices[
+ self.input_devices.index(input_device)
+ ]
+ sd.default.device[1] = self.output_devices_indices[
+ self.output_devices.index(output_device)
+ ]
+ printt("Input device: %s:%s", str(sd.default.device[0]), input_device)
+ printt("Output device: %s:%s", str(sd.default.device[1]), output_device)
+
+ def get_device_samplerate(self):
+ return int(
+ sd.query_devices(device=sd.default.device[0])["default_samplerate"]
+ )
+
+ def get_device_channels(self):
+ max_input_channels = sd.query_devices(device=sd.default.device[0])[
+ "max_input_channels"
+ ]
+ max_output_channels = sd.query_devices(device=sd.default.device[1])[
+ "max_output_channels"
+ ]
+ return min(max_input_channels, max_output_channels, 2)
+
+ gui = GUI()
diff --git a/i18n/i18n.py b/i18n/i18n.py
new file mode 100644
index 0000000000000000000000000000000000000000..ea555da8072bdf700d5b5b35c1a357dc724eb326
--- /dev/null
+++ b/i18n/i18n.py
@@ -0,0 +1,29 @@
+import json
+import locale
+import os
+from configs import singleton_variable
+
+
+def load_language_list(language):
+ with open(f"./i18n/locale/{language}.json", "r", encoding="utf-8") as f:
+ language_list = json.load(f)
+ return language_list
+
+
+@singleton_variable
+class I18nAuto:
+ def __init__(self, language=None):
+ if language in ["Auto", None]:
+ language = locale.getdefaultlocale(
+ envvars=("LANG", "LC_ALL", "LC_CTYPE", "LANGUAGE")
+ )[0]
+ if not os.path.exists(f"./i18n/locale/{language}.json"):
+ language = "en_US"
+ self.language = language
+ self.language_map = load_language_list(language)
+
+ def __call__(self, key):
+ return self.language_map.get(key, key)
+
+ def __repr__(self):
+ return "Language: " + self.language
diff --git a/i18n/locale/en_US.json b/i18n/locale/en_US.json
new file mode 100644
index 0000000000000000000000000000000000000000..82ed6d5fb3d9c21eea7b25802d23f4a44ec174b0
--- /dev/null
+++ b/i18n/locale/en_US.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### View model information\n> Only supported for small model files extracted from the 'weights' folder.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume",
+ "Algorithmic delays (ms)": "Algorithmic delays (ms)",
+ "All processes have been completed!": "All processes have been completed!",
+ "Audio device": "Audio device",
+ "Auto-detect index path and select from the dropdown": "Auto-detect index path and select from the dropdown",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').",
+ "Batch inference": "Batch inference",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.",
+ "Batch size per GPU": "Batch size per GPU",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement",
+ "Calculate": "Calculate",
+ "Choose sample rate of the device": "Choose sample rate of the device",
+ "Choose sample rate of the model": "Choose sample rate of the model",
+ "Convert": "Convert",
+ "Device type": "Device type",
+ "Enable phase vocoder": "Enable phase vocoder",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2",
+ "Enter the experiment name": "Enter the experiment name",
+ "Enter the path of the audio folder to be processed": "Enter the path of the audio folder to be processed",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)",
+ "Enter the path of the training folder": "Enter the path of the training folder",
+ "Exist": "Exist",
+ "Export Onnx": "Export Onnx",
+ "Export Onnx Model": "Export Onnx Model",
+ "Export audio (click on the three dots in the lower right corner to download)": "Export audio (click on the three dots in the lower right corner to download)",
+ "Export file format": "Export file format",
+ "Extra inference time": "Extra inference time",
+ "Extract": "Extract",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation",
+ "FAQ (Frequently Asked Questions)": "FAQ (Frequently Asked Questions)",
+ "Fade length": "Fade length",
+ "Fail": "Fail",
+ "Feature extraction": "Feature extraction",
+ "Feature searching ratio": "Feature searching ratio",
+ "Formant offset": "Formant offset",
+ "Fusion": "Fusion",
+ "GPU Information": "GPU Information",
+ "General settings": "General settings",
+ "Hidden": "Hidden",
+ "ID of model A (long)": "ID of model A (long)",
+ "ID of model B (long)": "ID of model B (long)",
+ "ID(long)": "ID(long)",
+ "ID(short)": "ID(short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.",
+ "Inference time (ms)": "Inference time (ms)",
+ "Inferencing voice": "Inferencing voice",
+ "Information": "Information",
+ "Input device": "Input device",
+ "Input noise reduction": "Input noise reduction",
+ "Input voice monitor": "Input voice monitor",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "Load model",
+ "Load pre-trained base model D path": "Load pre-trained base model D path",
+ "Load pre-trained base model G path": "Load pre-trained base model G path",
+ "Loudness factor": "Loudness factor",
+ "Model": "Model",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "Model Inference",
+ "Model architecture version": "Model architecture version",
+ "Model info": "Model info",
+ "Model information to be modified": "Model information to be modified",
+ "Model information to be placed": "Model information to be placed",
+ "Model name": "Model name",
+ "Modify": "Modify",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Multiple audio files can also be imported. If a folder path exists, this input is ignored.",
+ "No": "No",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "Number of CPU processes used for harvest pitch algorithm",
+ "Number of CPU processes used for pitch extraction and data processing": "Number of CPU processes used for pitch extraction and data processing",
+ "One-click training": "One-click training",
+ "Onnx Export Path": "Onnx Export Path",
+ "Output converted voice": "Output converted voice",
+ "Output device": "Output device",
+ "Output information": "Output information",
+ "Output noise reduction": "Output noise reduction",
+ "Path to Model": "Path to Model",
+ "Path to Model A": "Path to Model A",
+ "Path to Model B": "Path to Model B",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Path to the feature index file. Leave blank to use the selected result from the dropdown",
+ "Performance settings": "Performance settings",
+ "Pitch detection algorithm": "Pitch detection algorithm",
+ "Pitch guidance (f0)": "Pitch guidance (f0)",
+ "Pitch settings": "Pitch settings",
+ "Please choose the .index file": "Please choose the .index file",
+ "Please choose the .pth file": "Please choose the .pth file",
+ "Please specify the speaker/singer ID": "Please specify the speaker/singer ID",
+ "Process data": "Process data",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy",
+ "RVC Model Path": "RVC Model Path",
+ "Read from model": "Read from model",
+ "Refresh voice list and index path": "Refresh voice list and index path",
+ "Reload device list": "Reload device list",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling",
+ "Response threshold": "Response threshold",
+ "Sample length": "Sample length",
+ "Sampling rate": "Sampling rate",
+ "Save a small final model to the 'weights' folder at each save point": "Save a small final model to the 'weights' folder at each save point",
+ "Save file name (default: same as the source file)": "Save file name (default: same as the source file)",
+ "Save frequency (save_every_epoch)": "Save frequency (save_every_epoch)",
+ "Save name": "Save name",
+ "Save only the latest '.ckpt' file to save disk space": "Save only the latest '.ckpt' file to save disk space",
+ "Saved model name (without extension)": "Saved model name (without extension)",
+ "Sealing date": "Sealing date",
+ "Select Speaker/Singer ID": "Select Speaker/Singer ID",
+ "Select the .index file": "Select the .index file",
+ "Select the .pth file": "Select the .pth file",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "Single inference",
+ "Specify output folder": "Specify output folder",
+ "Specify the output folder for accompaniment": "Specify the output folder for accompaniment",
+ "Specify the output folder for vocals": "Specify the output folder for vocals",
+ "Start audio conversion": "Start audio conversion",
+ "Step 1: Processing data": "Step 1: Processing data",
+ "Step 3a: Model training started": "Step 3a: Model training started",
+ "Stop audio conversion": "Stop audio conversion",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "Target sample rate",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.",
+ "Total training epochs (total_epoch)": "Total training epochs (total_epoch)",
+ "Train": "Train",
+ "Train feature index": "Train feature index",
+ "Train model": "Train model",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Unfortunately, there is no compatible GPU available to support your training.",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "Unload model to save GPU memory",
+ "Version": "Version",
+ "View": "View",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Vocals/Accompaniment Separation & Reverberation Removal",
+ "Weight (w) for Model A": "Weight (w) for Model A",
+ "Whether the model has pitch guidance": "Whether the model has pitch guidance",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Whether the model has pitch guidance (1: yes, 0: no)",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Whether the model has pitch guidance (required for singing, optional for speech)",
+ "Yes": "Yes",
+ "ckpt Processing": "ckpt Processing",
+ "index path cannot contain unicode characters": "index path cannot contain unicode characters",
+ "pth path cannot contain unicode characters": "pth path cannot contain unicode characters",
+ "step2:Pitch extraction & feature extraction": "step2:Pitch extraction & feature extraction"
+}
diff --git a/i18n/locale/es_ES.json b/i18n/locale/es_ES.json
new file mode 100644
index 0000000000000000000000000000000000000000..fd84d4ab31db215e6b2571ed0548cda7f417851b
--- /dev/null
+++ b/i18n/locale/es_ES.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Comparación de modelos\n> Obtén el ID del modelo (largo) en la sección de `Ver información del modelo` a continuación\n\nSe puede utilizar para comparar la similitud de la inferencia de dos modelos.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Extracción de modelo\n> Ingrese la ruta de un archivo de modelo grande en la carpeta 'logs'.\n\nAplicable cuando desea extraer un archivo de modelo pequeño después de entrenar a mitad de camino y no se guardó automáticamente, o cuando desea probar un modelo intermedio.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Fusión de modelos\nSe puede utilizar para fusionar diferentes voces.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Modificar la información del modelo\n> Solo admite archivos de modelos pequeños extraídos en la carpeta 'weights'.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Paso 1. Complete la configuración del experimento.\nLos datos del experimento se almacenan en el directorio 'logs', con cada experimento en una carpeta separada. La ruta del nombre del experimento debe ingresarse manualmente y debe contener la configuración del experimento, los registros y los archivos del modelo entrenado.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Paso dos: Procesamiento de audio\n#### 1. Segmentación de audio\nRecorre automáticamente todos los archivos que se pueden decodificar en audio en la carpeta de entrenamiento y realiza la segmentación y normalización, generando 2 carpetas wav en el directorio del experimento; por ahora solo se admite el entrenamiento individual.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Paso tres: Comienza el entrenamiento\nCompleta la configuración de entrenamiento, comienza a entrenar el modelo y el índice.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Ver información del modelo\n> Solo aplicable a archivos de modelos pequeños extraídos de la carpeta 'weights'.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Extracción de características\nUtiliza la CPU para extraer el tono (si el modelo tiene tono), utiliza la GPU para extraer características (selecciona el número de tarjeta).",
+ "Actually calculated": "Valor realmente calculado",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Proporción de fusión para reemplazar el sobre de volumen de entrada con el sobre de volumen de salida, cuanto más cerca de 1, más se utiliza el sobre de salida",
+ "Algorithmic delays (ms)": "Retrasos algorítmicos (ms)",
+ "All processes have been completed!": "¡Todo el proceso ha terminado!",
+ "Audio device": "Dispositivo de audio",
+ "Auto-detect index path and select from the dropdown": "Detección automática de la ruta del índice, selección desplegable (dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Conversión por lotes, ingrese la carpeta que contiene los archivos de audio para convertir o cargue varios archivos de audio. El audio convertido se emitirá en la carpeta especificada (opción predeterminada).",
+ "Batch inference": "Inferencia por lotes",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Procesamiento por lotes para la separación de acompañamiento vocal utilizando el modelo UVR5.
Ejemplo de formato de ruta de carpeta válido: D:\\ruta\\a\\la\\carpeta\\de\\entrada (copiar desde la barra de direcciones del administrador de archivos).
El modelo se divide en tres categorías:
1. Preservar voces: Elija esta opción para audio sin armonías. Preserva las voces mejor que HP5. Incluye dos modelos incorporados: HP2 y HP3. HP3 puede filtrar ligeramente el acompañamiento pero conserva las voces un poco mejor que HP2.
2. Preservar solo voces principales: Elija esta opción para audio con armonías. Puede debilitar las voces principales. Incluye un modelo incorporado: HP5.
3. Modelos de des-reverberación y des-retardo (por FoxJoy):
(1) MDX-Net: La mejor opción para la eliminación de reverberación estéreo pero no puede eliminar la reverberación mono;
(234) DeEcho: Elimina efectos de retardo. El modo Agresivo elimina más a fondo que el modo Normal. DeReverb adicionalmente elimina la reverberación y puede eliminar la reverberación mono, pero no muy efectivamente para contenido de alta frecuencia fuertemente reverberado.
Notas de des-reverberación/des-retardo:
1. El tiempo de procesamiento para el modelo DeEcho-DeReverb es aproximadamente el doble que los otros dos modelos DeEcho.
2. El modelo MDX-Net-Dereverb es bastante lento.
3. La configuración más limpia recomendada es aplicar primero MDX-Net y luego DeEcho-Agresivo.",
+ "Batch size per GPU": "Tamaño del lote (batch_size) por tarjeta gráfica",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Si almacenar en caché todos los conjuntos de entrenamiento en la memoria de la GPU. Los conjuntos de datos pequeños (menos de 10 minutos) se pueden almacenar en caché para acelerar el entrenamiento, pero el almacenamiento en caché de conjuntos de datos grandes puede causar errores de memoria en la GPU y no aumenta la velocidad de manera significativa.",
+ "Calculate": "Calcularlo",
+ "Choose sample rate of the device": "Elija la frecuencia de muestreo del dispositivo",
+ "Choose sample rate of the model": "Elija la frecuencia de muestreo del modelo",
+ "Convert": "Conversión",
+ "Device type": "Tipo de dispositivo",
+ "Enable phase vocoder": "Habilitar vocoder de fase",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Separe los números de identificación de la GPU con '-' al ingresarlos. Por ejemplo, '0-1-2' significa usar GPU 0, GPU 1 y GPU 2.",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Separe los números de identificación de la GPU con '-' al ingresarlos. Por ejemplo, '0-1-2' significa usar GPU 0, GPU 1 y GPU 2.",
+ "Enter the experiment name": "Ingrese el nombre del modelo",
+ "Enter the path of the audio folder to be processed": "Ingrese la ruta a la carpeta de audio que se procesará",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Ingrese la ruta a la carpeta de audio que se procesará (simplemente cópiela desde la barra de direcciones del administrador de archivos)",
+ "Enter the path of the training folder": "Introduzca la ruta de la carpeta de entrenamiento",
+ "Exist": "Existente",
+ "Export Onnx": "Exportar Onnx",
+ "Export Onnx Model": "Exportar modelo Onnx",
+ "Export audio (click on the three dots in the lower right corner to download)": "Salida de audio (haga clic en los tres puntos en la esquina inferior derecha para descargar)",
+ "Export file format": "Formato de archivo de exportación",
+ "Extra inference time": "Tiempo de inferencia adicional",
+ "Extract": "Extraer",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "Archivo de curva F0, opcional, un tono por línea, en lugar de F0 predeterminado y cambio de tono",
+ "FAQ (Frequently Asked Questions)": "Preguntas frecuentes",
+ "Fade length": "Duración del fundido de entrada/salida",
+ "Fail": "Falló",
+ "Feature extraction": "Extracción de características",
+ "Feature searching ratio": "Proporción de función de búsqueda",
+ "Formant offset": "Compensación resonante",
+ "Fusion": "Fusión",
+ "GPU Information": "información de la GPU",
+ "General settings": "Configuración general",
+ "Hidden": "Oculto",
+ "ID of model A (long)": "ID del modelo A (largo)",
+ "ID of model B (long)": "ID del modelo B (largo)",
+ "ID(long)": "ID (largo)",
+ "ID(short)": "ID (corto)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "Si es >=3, entonces use el resultado del reconocimiento de tono de 'harvest' con filtro de mediana, el valor es el radio del filtro, su uso puede debilitar el sonido sordo",
+ "Inference time (ms)": "Inferir tiempo (ms)",
+ "Inferencing voice": "inferencia de voz",
+ "Information": "Información",
+ "Input device": "Dispositivo de entrada",
+ "Input noise reduction": "Reducción de ruido de entrada",
+ "Input voice monitor": "Monitor de voz de entrada",
+ "Link index to outside folder": "Vincular índice a carpeta externa",
+ "Load model": "Cargar modelo",
+ "Load pre-trained base model D path": "Cargue la ruta del modelo D base pre-entrenada.",
+ "Load pre-trained base model G path": "Cargue la ruta del modelo G base pre-entrenada.",
+ "Loudness factor": "factor de sonoridad",
+ "Model": "Modelo",
+ "Model Author": "Autor del modelo",
+ "Model Author (Nullable)": "Autor del modelo (anulable)",
+ "Model Inference": "inferencia del modelo",
+ "Model architecture version": "Versión y modelo del modelo",
+ "Model info": "Información del modelo",
+ "Model information to be modified": "Información del modelo a modificar",
+ "Model information to be placed": "Información del modelo a colocar.",
+ "Model name": "Nombre del modelo",
+ "Modify": "Modificar",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "También se pueden importar varios archivos de audio. Si existe una ruta de carpeta, esta entrada se ignora.",
+ "No": "No",
+ "None": "Ninguno",
+ "Not exist": "Inexistente",
+ "Number of CPU processes used for harvest pitch algorithm": "Número de procesos",
+ "Number of CPU processes used for pitch extraction and data processing": "Número de procesos de CPU utilizados para extraer el tono y procesar los datos",
+ "One-click training": "Entrenamiento con un clic",
+ "Onnx Export Path": "Ruta de salida Onnx",
+ "Output converted voice": "Salida de voz convertida",
+ "Output device": "Dispositivo de salida",
+ "Output information": "Información de salida",
+ "Output noise reduction": "Reducción de ruido de salida",
+ "Path to Model": "Ruta del modelo",
+ "Path to Model A": "Modelo A ruta.",
+ "Path to Model B": "Modelo B ruta.",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Ruta del archivo de la biblioteca de características, si está vacío, se utilizará el resultado de la selección desplegable",
+ "Performance settings": "Configuración de rendimiento",
+ "Pitch detection algorithm": "Algoritmo de tono",
+ "Pitch guidance (f0)": "Guía de tono (f0)",
+ "Pitch settings": "Ajuste de tono",
+ "Please choose the .index file": "Seleccione el archivo .index",
+ "Please choose the .pth file": "Seleccione el archivo .pth",
+ "Please specify the speaker/singer ID": "ID del modelo",
+ "Process data": "Procesar datos",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Proteger las consonantes claras y la respiración, prevenir artefactos como la distorsión de sonido electrónico, 0.5 no está activado, reducir aumentará la protección pero puede reducir el efecto del índice",
+ "RVC Model Path": "Ruta del modelo RVC",
+ "Read from model": "Leer del modelo",
+ "Refresh voice list and index path": "Actualizar la lista de modelos e índice de rutas",
+ "Reload device list": "Actualizar lista de dispositivos",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Remuestreo posterior al proceso a la tasa de muestreo final, 0 significa no remuestrear",
+ "Response threshold": "Umbral de respuesta",
+ "Sample length": "Longitud de muestreo",
+ "Sampling rate": "Tasa de muestreo",
+ "Save a small final model to the 'weights' folder at each save point": "Guardar pequeño modelo final en la carpeta 'weights' en cada punto de guardado",
+ "Save file name (default: same as the source file)": "Nombre del archivo que se guardará, el valor predeterminado es el mismo que el nombre del archivo de origen",
+ "Save frequency (save_every_epoch)": "Frecuencia de guardado (save_every_epoch)",
+ "Save name": "Guardar nombre",
+ "Save only the latest '.ckpt' file to save disk space": "Guardar solo el archivo ckpt más reciente para ahorrar espacio en disco",
+ "Saved model name (without extension)": "Nombre del modelo guardado sin extensión.",
+ "Sealing date": "fecha de sellado",
+ "Select Speaker/Singer ID": "Seleccione una identificación de altavoz",
+ "Select the .index file": "Seleccione el archivo .index",
+ "Select the .pth file": "Seleccione el archivo .pth",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Seleccione el algoritmo de extracción de tono, use 'pm' para acelerar la entrada de canto, 'harvest' es bueno para los graves pero extremadamente lento, 'crepe' tiene buenos resultados pero consume GPU",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Seleccione el algoritmo de extracción de tono: la canción de entrada se puede acelerar con pm, la voz de alta calidad pero CPU pobre se puede acelerar con dio, harvest es mejor pero más lento, rmvpe es el mejor y se come ligeramente la CPU/GPU",
+ "Similarity": "Semejanza",
+ "Similarity (from 0 to 1)": "Similitud (de 0 a 1)",
+ "Single inference": "Inferencia única",
+ "Specify output folder": "Especificar carpeta de salida",
+ "Specify the output folder for accompaniment": "Especifique la carpeta de salida para las voces no principales",
+ "Specify the output folder for vocals": "Especifique la carpeta de salida para la voz principal",
+ "Start audio conversion": "Iniciar conversión de audio",
+ "Step 1: Processing data": "Paso 1: Procesando datos",
+ "Step 3a: Model training started": "Paso 3a: Entrenando el modelo",
+ "Stop audio conversion": "Detener la conversión de audio",
+ "Successfully built index into": "Índice construido con éxito en",
+ "Takeover WASAPI device": "Adquisición del dispositivo WASAPI",
+ "Target sample rate": "Tasa de muestreo objetivo",
+ "The audio file to be processed": "El archivo de audio a procesar",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "Este software es de código abierto bajo la licencia MIT, el autor no tiene ningún control sobre el software, y aquellos que usan el software y difunden los sonidos exportados por el software son los únicos responsables.
Si no está de acuerdo con esta cláusula , no puede utilizar ni citar ningún código ni archivo del paquete de software Consulte el directorio raíz Agreement-LICENSE.txt para obtener más información.",
+ "Total training epochs (total_epoch)": "Total de épocas de entrenamiento (total_epoch)",
+ "Train": "Entrenamiento",
+ "Train feature index": "Índice de características",
+ "Train model": "Entrenar Modelo",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Entrenamiento finalizado, puede ver el registro de entrenamiento en la consola o en el archivo train.log en la carpeta del experimento",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Cambio de tono (entero, número de semitonos, subir una octava +12 o bajar una octava -12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Lamentablemente, no tiene una tarjeta gráfica adecuada para soportar su entrenamiento",
+ "Unknown": "Desconocido",
+ "Unload model to save GPU memory": "Descargue la voz para ahorrar memoria GPU",
+ "Version": "Versión",
+ "View": "Ver",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Separación de voz acompañante & eliminación de reverberación & eco",
+ "Weight (w) for Model A": "Un peso modelo para el modelo A.",
+ "Whether the model has pitch guidance": "Si el modelo tiene guía de tono.",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Si el modelo tiene guía de tono, 1 para sí, 0 para no",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Si el modelo tiene guía de tono (necesaria para cantar, pero no para hablar)",
+ "Yes": "Sí",
+ "ckpt Processing": "Procesamiento de recibos",
+ "index path cannot contain unicode characters": "La ruta del archivo .index no debe contener caracteres chinos.",
+ "pth path cannot contain unicode characters": "La ruta del archivo .pth no debe contener caracteres chinos.",
+ "step2:Pitch extraction & feature extraction": "Paso 2: Extracción del tono y extracción de características"
+}
diff --git a/i18n/locale/fr_FR.json b/i18n/locale/fr_FR.json
new file mode 100644
index 0000000000000000000000000000000000000000..ad9bce588dd51f0c11c6467a3fc25533668a5861
--- /dev/null
+++ b/i18n/locale/fr_FR.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Comparaison des modèles\n> Pour obtenir l'ID du modèle (long), veuillez consulter la section `Voir les informations du modèle` ci-dessous.\n\nPeut être utilisé pour comparer la similarité des inférences entre deux modèles.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Extraction du modèle\n> Saisissez le chemin d'accès au modèle du grand fichier dans le dossier \"logs\".\n\nCette fonction est utile si vous souhaitez arrêter l'entrainement à mi-chemin et extraire et enregistrer manuellement un petit fichier de modèle, ou si vous souhaitez tester un modèle intermédiaire.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Modifier les informations du modèle\n> Uniquement pris en charge pour les petits fichiers de modèle extraits du dossier 'weights'.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Étape 1. Remplissez la configuration expérimentale.\nLes données expérimentales sont stockées dans le dossier 'logs', avec chaque expérience ayant un dossier distinct. Entrez manuellement le chemin du nom de l'expérience, qui contient la configuration expérimentale, les journaux et les fichiers de modèle entraînés.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Deuxième étape : Traitement audio\n#### 1. Découpage de l'audio\nParcourez automatiquement tous les fichiers qui peuvent être décodés en audio dans le dossier d'entraînement et effectuez le découpage et la normalisation. Deux dossiers wav sont générés dans le répertoire de l'expérience. Pour le moment, seul l'entraînement individuel est pris en charge.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Troisième étape : Commencer l'entraînement\nRemplissez les paramètres d'entraînement, commencez à entraîner le modèle et l'index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Afficher les informations sur le modèle\n> Uniquement pour les petits fichiers de modèle extraits du dossier 'weights'.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Extraction des caractéristiques\nUtilisez le CPU pour extraire la hauteur tonale (si le modèle a une hauteur tonale), utilisez le GPU pour extraire les caractéristiques (sélectionnez le numéro de carte).",
+ "Actually calculated": "Effectivement calculé",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Ajustez l'échelle de l'enveloppe de volume. Plus il est proche de 0, plus il imite le volume des voix originales. Cela peut aider à masquer les bruits et à rendre le volume plus naturel lorsqu'il est réglé relativement bas. Plus le volume est proche de 1, plus le volume sera fort et constant :",
+ "Algorithmic delays (ms)": "Délais algorithmiques (ms)",
+ "All processes have been completed!": "Toutes les étapes ont été terminées !",
+ "Audio device": "Périphérique audio",
+ "Auto-detect index path and select from the dropdown": "Détecter automatiquement le chemin d'accès à l'index et le sélectionner dans la liste déroulante :",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Conversion en lot. Entrez le dossier contenant les fichiers audio à convertir ou téléchargez plusieurs fichiers audio. Les fichiers audio convertis seront enregistrés dans le dossier spécifié (par défaut : 'opt').",
+ "Batch inference": "Inférence par lots",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Traitement en lot pour la séparation de la voix et de l'accompagnement vocal à l'aide du modèle UVR5.
Exemple d'un format de chemin de dossier valide : D:\\chemin\\vers\\dossier\\d'entrée (copiez-le depuis la barre d'adresse du gestionnaire de fichiers).
Le modèle est divisé en trois catégories :
1. Préserver la voix : Choisissez cette option pour l'audio sans harmonies. Elle préserve la voix mieux que HP5. Il comprend deux modèles intégrés : HP2 et HP3. HP3 peut légèrement laisser passer l'accompagnement mais préserve légèrement mieux la voix que HP2.
2. Préserver uniquement la voix principale : Choisissez cette option pour l'audio avec harmonies. Cela peut affaiblir la voix principale. Il comprend un modèle intégré : HP5.
3. Modèles de suppression de la réverbération et du délai (par FoxJoy) :
(1) MDX-Net : Le meilleur choix pour la suppression de la réverbération stéréo, mais ne peut pas supprimer la réverbération mono.
(234) DeEcho : Supprime les effets de délai. Le mode Aggressive supprime plus efficacement que le mode Normal. DeReverb supprime également la réverbération et peut supprimer la réverbération mono, mais pas très efficacement pour les contenus à haute fréquence fortement réverbérés.
Notes sur la suppression de la réverbération et du délai :
1. Le temps de traitement pour le modèle DeEcho-DeReverb est environ deux fois plus long que pour les autres deux modèles DeEcho.
2. Le modèle MDX-Net-Dereverb est assez lent.
3. La configuration la plus propre recommandée est d'appliquer d'abord MDX-Net, puis DeEcho-Aggressive.",
+ "Batch size per GPU": "Taille du batch par GPU :",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Mettre en cache tous les ensembles d'entrainement dans la mémoire GPU. Mettre en cache de petits ensembles de données (moins de 10 minutes) peut accélérer l'entrainement, mais mettre en cache de grands ensembles de données consommera beaucoup de mémoire GPU et peut ne pas apporter beaucoup d'amélioration de vitesse :",
+ "Calculate": "Calculez-le",
+ "Choose sample rate of the device": "Choisissez la fréquence d'échantillonnage de l'appareil",
+ "Choose sample rate of the model": "Choisissez la fréquence d'échantillonnage du modèle",
+ "Convert": "Convertir",
+ "Device type": "Type d'appareil",
+ "Enable phase vocoder": "Activer le vocodeur de phase",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Configuration des numéros de carte RMVPE : séparez les index GPU par des tirets \"-\", par exemple, 0-0-1 pour utiliser 2 processus sur GPU0 et 1 processus sur GPU1.",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Entrez le(s) index GPU séparé(s) par '-', par exemple, 0-1-2 pour utiliser les GPU 0, 1 et 2 :",
+ "Enter the experiment name": "Saisissez le nom de l'expérience :",
+ "Enter the path of the audio folder to be processed": "Entrez le chemin du dossier audio à traiter :",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Entrez le chemin du dossier audio à traiter (copiez-le depuis la barre d'adresse du gestionnaire de fichiers) :",
+ "Enter the path of the training folder": "Indiquez le chemin d'accès au dossier d'entraînement :",
+ "Exist": "Existe",
+ "Export Onnx": "Exporter en ONNX",
+ "Export Onnx Model": "Exporter le modèle au format ONNX.",
+ "Export audio (click on the three dots in the lower right corner to download)": "Exporter l'audio (cliquer sur les trois points dans le coin inférieur droit pour télécharger)",
+ "Export file format": "Format de fichier d'exportation",
+ "Extra inference time": "Temps d'inférence supplémentaire",
+ "Extract": "Extraire",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "Fichier de courbe F0 (facultatif). Une hauteur par ligne. Remplace la fréquence fondamentale par défaut et la modulation de la hauteur :",
+ "FAQ (Frequently Asked Questions)": "FAQ (Foire Aux Questions)",
+ "Fade length": "Longueur de la transition",
+ "Fail": "Raté",
+ "Feature extraction": "Extraction des caractéristiques",
+ "Feature searching ratio": "Rapport de recherche de caractéristiques (contrôle l'intensité de l'accent, un rapport trop élevé provoque des artefacts) :",
+ "Formant offset": "Décalage résonant",
+ "Fusion": "Fusion",
+ "GPU Information": "Informations sur la carte graphique (GPU)",
+ "General settings": "Paramètres généraux",
+ "Hidden": "Caché",
+ "ID of model A (long)": "ID du modèle A (long)",
+ "ID of model B (long)": "ID du modèle B (long)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (court)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "Si >=3 : appliquer un filtrage médian aux résultats de la reconnaissance de la hauteur de récolte. La valeur représente le rayon du filtre et peut réduire la respiration.",
+ "Inference time (ms)": "Temps d'inférence (ms)",
+ "Inferencing voice": "Voix pour l'inférence",
+ "Information": "Information",
+ "Input device": "Dispositif d'entrée",
+ "Input noise reduction": "Réduction du bruit d'entrée",
+ "Input voice monitor": "Moniteur vocal d'entrée",
+ "Link index to outside folder": "Lier l'index au dossier extérieur",
+ "Load model": "Charger le modèle.",
+ "Load pre-trained base model D path": "Charger le chemin du modèle de base pré-entraîné D :",
+ "Load pre-trained base model G path": "Charger le chemin du modèle de base pré-entraîné G :",
+ "Loudness factor": "Facteur de volume sonore",
+ "Model": "Modèle",
+ "Model Author": "Auteur du modèle",
+ "Model Author (Nullable)": "Auteur du modèle (Nullable)",
+ "Model Inference": "Inférence du modèle",
+ "Model architecture version": "Version de l'architecture du modèle :",
+ "Model info": "Informations sur le modèle",
+ "Model information to be modified": "Informations sur le modèle à modifier :",
+ "Model information to be placed": "Informations sur le modèle à placer :",
+ "Model name": "Nom du modèle",
+ "Modify": "Modifier",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Il est également possible d'importer plusieurs fichiers audio. Si un chemin de dossier existe, cette entrée est ignorée.",
+ "No": "Non",
+ "None": "Aucun",
+ "Not exist": "N'existe",
+ "Number of CPU processes used for harvest pitch algorithm": "Nombre de processus CPU utilisés pour l'algorithme de reconnaissance de la hauteur (pitch) dans le cadre de la récolte (harvest).",
+ "Number of CPU processes used for pitch extraction and data processing": "Nombre de processus CPU utilisés pour l'extraction de la hauteur et le traitement des données :",
+ "One-click training": "Entraînement en un clic",
+ "Onnx Export Path": "Chemin d'exportation ONNX :",
+ "Output converted voice": "Sortie voix convertie",
+ "Output device": "Dispositif de sortie",
+ "Output information": "Informations sur la sortie",
+ "Output noise reduction": "Réduction du bruit de sortie",
+ "Path to Model": "Le chemin vers le modèle :",
+ "Path to Model A": "Chemin d'accès au modèle A :",
+ "Path to Model B": "Chemin d'accès au modèle B :",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Chemin d'accès au fichier d'index des caractéristiques. Laisser vide pour utiliser le résultat sélectionné dans la liste déroulante :",
+ "Performance settings": "Paramètres de performance",
+ "Pitch detection algorithm": "Algorithme de détection de la hauteur",
+ "Pitch guidance (f0)": "Guidage du pas (f0)",
+ "Pitch settings": "Réglages de la hauteur",
+ "Please choose the .index file": "Veuillez sélectionner le fichier d'index",
+ "Please choose the .pth file": "Veuillez sélectionner le fichier pth",
+ "Please specify the speaker/singer ID": "Veuillez spécifier l'ID de l'orateur ou du chanteur :",
+ "Process data": "Traitement des données",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Protéger les consonnes sourdes et les bruits de respiration pour éviter les artefacts tels que le déchirement dans la musique électronique. Réglez à 0,5 pour désactiver. Diminuez la valeur pour renforcer la protection, mais cela peut réduire la précision de l'indexation :",
+ "RVC Model Path": "Chemin du modèle RVC :",
+ "Read from model": "Lire à partir du modèle",
+ "Refresh voice list and index path": "Actualiser la liste des voix et le vers l'index.",
+ "Reload device list": "Recharger la liste des dispositifs",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Rééchantillonner l'audio de sortie en post-traitement à la fréquence d'échantillonnage finale. Réglez sur 0 pour ne pas effectuer de rééchantillonnage :",
+ "Response threshold": "Seuil de réponse",
+ "Sample length": "Longueur de l'échantillon",
+ "Sampling rate": "Taux d'échantillonnage",
+ "Save a small final model to the 'weights' folder at each save point": "Enregistrer un petit modèle final dans le dossier 'weights' à chaque point de sauvegarde :",
+ "Save file name (default: same as the source file)": "Nom du fichier de sauvegarde (par défaut : identique au nom du fichier source) :",
+ "Save frequency (save_every_epoch)": "Fréquence de sauvegarde (save_every_epoch) :",
+ "Save name": "Nom de sauvegarde :",
+ "Save only the latest '.ckpt' file to save disk space": "Enregistrer uniquement le dernier fichier '.ckpt' pour économiser de l'espace disque :",
+ "Saved model name (without extension)": "Nom du modèle enregistré (sans extension) :",
+ "Sealing date": "Date de scellement",
+ "Select Speaker/Singer ID": "Sélectionner l'ID de l'orateur ou du chanteur :",
+ "Select the .index file": "Sélectionner le fichier .index",
+ "Select the .pth file": "Sélectionner le fichier .pth",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Sélectionnez l'algorithme d'extraction de la hauteur de ton (\"pm\" : extraction plus rapide mais parole de moindre qualité ; \"harvest\" : meilleure basse mais extrêmement lente ; \"crepe\" : meilleure qualité mais utilisation intensive du GPU), \"rmvpe\" : meilleure qualité et peu d'utilisation du GPU.",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Sélection de l'algorithme d'extraction de la hauteur : la chanson d'entrée peut être traitée plus rapidement par pm, avec une voix de haute qualité mais un CPU médiocre, par dio, harvest est meilleur mais plus lent, rmvpe est le meilleur, mais consomme légèrement le CPU/GPU.",
+ "Similarity": "Similarité",
+ "Similarity (from 0 to 1)": "Similarité (de 0 à 1)",
+ "Single inference": "Inférence unique",
+ "Specify output folder": "Spécifiez le dossier de sortie :",
+ "Specify the output folder for accompaniment": "Spécifiez le dossier de sortie pour l'accompagnement :",
+ "Specify the output folder for vocals": "Spécifiez le dossier de sortie pour les fichiers de voix :",
+ "Start audio conversion": "Démarrer la conversion audio.",
+ "Step 1: Processing data": "Étape 1 : Traitement des données en cours.",
+ "Step 3a: Model training started": "Étape 3a : L'entraînement du modèle a commencé.",
+ "Stop audio conversion": "Arrêter la conversion audio",
+ "Successfully built index into": "Index intégré avec succès dans",
+ "Takeover WASAPI device": "Reprise du périphérique WASAPI",
+ "Target sample rate": "Taux d'échantillonnage cible :",
+ "The audio file to be processed": "Le fichier audio à traiter",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "Ce logiciel est open source sous la licence MIT. L'auteur n'a aucun contrôle sur le logiciel. Les utilisateurs qui utilisent le logiciel et distribuent les sons exportés par le logiciel en sont entièrement responsables.
Si vous n'acceptez pas cette clause, vous ne pouvez pas utiliser ou faire référence à aucun code ni fichier contenu dans le package logiciel. Consultez le fichier Agreement-LICENSE.txt dans le répertoire racine pour plus de détails.",
+ "Total training epochs (total_epoch)": "Nombre total d'époques d'entraînement (total_epoch) :",
+ "Train": "Entraîner",
+ "Train feature index": "Entraîner l'index des caractéristiques",
+ "Train model": "Entraîner le modèle",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Entraînement terminé. Vous pouvez consulter les rapports d'entraînement dans la console ou dans le fichier 'train.log' situé dans le dossier de l'expérience.",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Transposer (entier, nombre de demi-tons, monter d'une octave : 12, descendre d'une octave : -12) :",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Malheureusement, il n'y a pas de GPU compatible disponible pour prendre en charge votre entrainement.",
+ "Unknown": "Inconnu",
+ "Unload model to save GPU memory": "Décharger la voix pour économiser la mémoire GPU.",
+ "Version": "Version",
+ "View": "Voir",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Séparation des voix/accompagnement et suppression de la réverbération",
+ "Weight (w) for Model A": "Poids (w) pour le modèle A :",
+ "Whether the model has pitch guidance": "Indique si le modèle dispose d'un guidage en hauteur :",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Le modèle dispose-t-il d'un guide de hauteur (1 : oui, 0 : non) ?",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Indique si le modèle dispose d'un système de guidage de la hauteur (obligatoire pour le chant, facultatif pour la parole) :",
+ "Yes": "Oui",
+ "ckpt Processing": "Traitement des fichiers .ckpt",
+ "index path cannot contain unicode characters": "Le chemin du fichier d'index ne doit pas contenir de caractères chinois.",
+ "pth path cannot contain unicode characters": "Le chemin du fichier .pth ne doit pas contenir de caractères chinois.",
+ "step2:Pitch extraction & feature extraction": "Étape 2 : Extraction de la hauteur et extraction des caractéristiques en cours."
+}
diff --git a/i18n/locale/it_IT.json b/i18n/locale/it_IT.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1d0037f09ca36c67af15494a082e7e6ff6a4fc5
--- /dev/null
+++ b/i18n/locale/it_IT.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Estrazione del modello\n> Inserire il percorso del modello di file di grandi dimensioni nella cartella \"logs\".",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nPuò essere utilizzato per testare la fusione timbrica.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Modifica le informazioni sul modello\n> Supportato solo per i file di modello di piccole dimensioni estratti dalla cartella 'weights'.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Visualizza le informazioni sul modello\n> Supportato solo per file di modello piccoli estratti dalla cartella 'weights'.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Regola il ridimensionamento dell'inviluppo del volume. ",
+ "Algorithmic delays (ms)": "Algorithmic delays (ms)",
+ "All processes have been completed!": "Tutti i processi sono stati completati!",
+ "Audio device": "Dispositivo audio",
+ "Auto-detect index path and select from the dropdown": "Rileva automaticamente il percorso dell'indice e seleziona dal menu a tendina:",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Conversione massiva. Inserisci il percorso della cartella che contiene i file da convertire o carica più file audio. I file convertiti finiranno nella cartella specificata. (default: opt) ",
+ "Batch inference": "Batch inference",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Elaborazione batch per la separazione dell'accompagnamento vocale utilizzando il modello UVR5.
Esempio di un formato di percorso di cartella valido: D:\\path\\to\\input\\folder (copialo dalla barra degli indirizzi del file manager).
Il modello è suddiviso in tre categorie:
1. Conserva la voce: scegli questa opzione per l'audio senza armonie.
2. Mantieni solo la voce principale: scegli questa opzione per l'audio con armonie.
3. Modelli di de-riverbero e de-delay (di FoxJoy):
(1) MDX-Net: la scelta migliore per la rimozione del riverbero stereo ma non può rimuovere il riverbero mono;
Note di de-riverbero/de-delay:
1. Il tempo di elaborazione per il modello DeEcho-DeReverb è circa il doppio rispetto agli altri due modelli DeEcho.
2. Il modello MDX-Net-Dereverb è piuttosto lento.
3. La configurazione più pulita consigliata consiste nell'applicare prima MDX-Net e poi DeEcho-Aggressive.",
+ "Batch size per GPU": "Dimensione batch per GPU:",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Memorizza nella cache tutti i set di addestramento nella memoria della GPU. ",
+ "Calculate": "Calculate",
+ "Choose sample rate of the device": "Choose sample rate of the device",
+ "Choose sample rate of the model": "Choose sample rate of the model",
+ "Convert": "Convertire",
+ "Device type": "Device type",
+ "Enable phase vocoder": "Enable phase vocoder",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Inserisci gli indici GPU separati da '-', ad esempio 0-1-2 per utilizzare GPU 0, 1 e 2:",
+ "Enter the experiment name": "Inserisci il nome dell'esperimento:",
+ "Enter the path of the audio folder to be processed": "Immettere il percorso della cartella audio da elaborare:",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Immettere il percorso della cartella audio da elaborare (copiarlo dalla barra degli indirizzi del file manager):",
+ "Enter the path of the training folder": "Inserisci il percorso della cartella di addestramento:",
+ "Exist": "Exist",
+ "Export Onnx": "Esporta Onnx",
+ "Export Onnx Model": "Esporta modello Onnx",
+ "Export audio (click on the three dots in the lower right corner to download)": "Esporta audio (clicca sui tre puntini in basso a destra per scaricarlo)",
+ "Export file format": "Formato file di esportazione",
+ "Extra inference time": "Tempo di inferenza extra",
+ "Extract": "Estrai",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "File curva F0 (opzionale). ",
+ "FAQ (Frequently Asked Questions)": "FAQ (Domande frequenti)",
+ "Fade length": "Lunghezza dissolvenza",
+ "Fail": "失败",
+ "Feature extraction": "Estrazione delle caratteristiche",
+ "Feature searching ratio": "Rapporto funzionalità di ricerca (controlla la forza dell'accento, troppo alto ha artefatti):",
+ "Formant offset": "Formant offset",
+ "Fusion": "Fusione",
+ "GPU Information": "Informazioni GPU",
+ "General settings": "Impostazioni generali",
+ "Hidden": "Hidden",
+ "ID of model A (long)": "ID of model A (long)",
+ "ID of model B (long)": "ID of model B (long)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "Se >=3: applica il filtro mediano ai risultati del pitch raccolto. ",
+ "Inference time (ms)": "Tempo di inferenza (ms)",
+ "Inferencing voice": "Voce di inferenza:",
+ "Information": "Information",
+ "Input device": "Dispositivo di input",
+ "Input noise reduction": "Riduzione del rumore in ingresso",
+ "Input voice monitor": "Input voice monitor",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "Carica modello",
+ "Load pre-trained base model D path": "Carica il percorso D del modello base pre-addestrato:",
+ "Load pre-trained base model G path": "Carica il percorso G del modello base pre-addestrato:",
+ "Loudness factor": "fattore di sonorità",
+ "Model": "Modello",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "Inferenza del modello",
+ "Model architecture version": "Versione dell'architettura del modello:",
+ "Model info": "Model info",
+ "Model information to be modified": "Informazioni sul modello da modificare:",
+ "Model information to be placed": "Informazioni sul modello da posizionare:",
+ "Model name": "Model name",
+ "Modify": "Modificare",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Multiple audio files can also be imported. If a folder path exists, this input is ignored.",
+ "No": "NO",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "Number of CPU processes used for harvest pitch algorithm",
+ "Number of CPU processes used for pitch extraction and data processing": "Numero di processi CPU utilizzati per l'estrazione del tono e l'elaborazione dei dati:",
+ "One-click training": "Addestramento con un clic",
+ "Onnx Export Path": "Percorso di esportazione Onnx:",
+ "Output converted voice": "Output converted voice",
+ "Output device": "Dispositivo di uscita",
+ "Output information": "Informazioni sull'uscita",
+ "Output noise reduction": "Riduzione del rumore in uscita",
+ "Path to Model": "Percorso al modello:",
+ "Path to Model A": "Percorso per il modello A:",
+ "Path to Model B": "Percorso per il modello B:",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Percorso del file di indice delle caratteristiche. ",
+ "Performance settings": "Impostazioni delle prestazioni",
+ "Pitch detection algorithm": "Pitch detection algorithm",
+ "Pitch guidance (f0)": "Pitch guidance (f0)",
+ "Pitch settings": "Impostazioni del tono",
+ "Please choose the .index file": "Please choose the .index file",
+ "Please choose the .pth file": "Please choose the .pth file",
+ "Please specify the speaker/singer ID": "Si prega di specificare l'ID del locutore/cantante:",
+ "Process data": "Processa dati",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Proteggi le consonanti senza voce e i suoni del respiro per evitare artefatti come il tearing nella musica elettronica. ",
+ "RVC Model Path": "Percorso modello RVC:",
+ "Read from model": "Read from model",
+ "Refresh voice list and index path": "Aggiorna l'elenco delle voci e il percorso dell'indice",
+ "Reload device list": "Ricaricare l'elenco dei dispositivi",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Ricampiona l'audio di output in post-elaborazione alla frequenza di campionamento finale. ",
+ "Response threshold": "Soglia di risposta",
+ "Sample length": "Lunghezza del campione",
+ "Sampling rate": "Sampling rate",
+ "Save a small final model to the 'weights' folder at each save point": "Salva un piccolo modello finale nella cartella \"weights\" in ogni punto di salvataggio:",
+ "Save file name (default: same as the source file)": "Salva il nome del file (predefinito: uguale al file di origine):",
+ "Save frequency (save_every_epoch)": "Frequenza di salvataggio (save_every_epoch):",
+ "Save name": "Salva nome:",
+ "Save only the latest '.ckpt' file to save disk space": "Salva solo l'ultimo file '.ckpt' per risparmiare spazio su disco:",
+ "Saved model name (without extension)": "Nome del modello salvato (senza estensione):",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "Seleziona ID locutore/cantante:",
+ "Select the .index file": "Seleziona il file .index",
+ "Select the .pth file": "Seleziona il file .pth",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Seleziona l'algoritmo di estrazione del tono (\"pm\": estrazione più veloce ma risultato di qualità inferiore; \"harvest\": bassi migliori ma estremamente lenti; \"crepe\": qualità migliore ma utilizzo intensivo della GPU):",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "Single inference",
+ "Specify output folder": "Specifica la cartella di output:",
+ "Specify the output folder for accompaniment": "Specificare la cartella di output per l'accompagnamento:",
+ "Specify the output folder for vocals": "Specifica la cartella di output per le voci:",
+ "Start audio conversion": "Avvia la conversione audio",
+ "Step 1: Processing data": "Passaggio 1: elaborazione dei dati",
+ "Step 3a: Model training started": "Passaggio 3a: è iniziato l'addestramento del modello",
+ "Stop audio conversion": "Arresta la conversione audio",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "Frequenza di campionamento target:",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "Questo software è open source con licenza MIT.
Se non si accetta questa clausola, non è possibile utilizzare o fare riferimento a codici e file all'interno del pacchetto software. Contratto-LICENZA.txt per dettagli.",
+ "Total training epochs (total_epoch)": "Epoch totali di addestramento (total_epoch):",
+ "Train": "Addestramento",
+ "Train feature index": "Addestra indice delle caratteristiche",
+ "Train model": "Addestra modello",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Addestramento completato. ",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Trasposizione (numero intero, numero di semitoni, alza di un'ottava: 12, abbassa di un'ottava: -12):",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Sfortunatamente, non è disponibile alcuna GPU compatibile per supportare l'addestramento.",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "Scarica la voce per risparmiare memoria della GPU:",
+ "Version": "Versione",
+ "View": "Visualizzazione",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Separazione voce/accompagnamento",
+ "Weight (w) for Model A": "Peso (w) per il modello A:",
+ "Whether the model has pitch guidance": "Se il modello ha una guida del tono:",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Se il modello ha una guida del tono (1: sì, 0: no):",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Se il modello ha una guida del tono (necessario per il canto, facoltativo per il parlato):",
+ "Yes": "SÌ",
+ "ckpt Processing": "Elaborazione ckpt",
+ "index path cannot contain unicode characters": "index path cannot contain unicode characters",
+ "pth path cannot contain unicode characters": "pth è un'app per il futuro",
+ "step2:Pitch extraction & feature extraction": "step2:Pitch extraction & feature extraction"
+}
diff --git a/i18n/locale/ja_JP.json b/i18n/locale/ja_JP.json
new file mode 100644
index 0000000000000000000000000000000000000000..2a4e4ca73dc50f7d76c6945d520912d09771bc37
--- /dev/null
+++ b/i18n/locale/ja_JP.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### モデル比べ\n> モデルID(長)は下の`モデル情報を表示`に得ることが出来ます。\n\n両モデルの推論相似度を比べることが出来ます。",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### モデル抽出\n> ログフォルダー内の大モデルのパスを入力\n\nモデルを半分まで学習し、小モデルを保存しなかった場合、又は中間モデルをテストしたい場合に適用されます。",
+ "### Model fusion\nCan be used to test timbre fusion.": "### モデルマージ\n音源のマージテストに使用できます",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### モデル情報の修正\n> `weights`フォルダから抽出された小モデルのみ対応",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### 第一歩 実験設定入力\n実験データはlogsフォルダーに、実験名別のフォルダで保存されたため、その実験名をご自分で決定する必要があります。実験設定、ログ、学習されたモデルファイルなどがそのフォルダに含まれています。",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### 第二歩 音声処理\n#### 1. 音声切分\n学習フォルダー内のすべての音声ファイルを自動的に探し出し、切分と正規化を行い、2つのwavフォルダーを実験ディレクトリに生成します。現在は単人モデルの学習のみを支援しています。",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### 第三歩 学習開始\n学習設定を入力して、モデルと索引の学習を開始します。",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### モデル情報を表示\n> `weights`フォルダから抽出された小さなのみ対応",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. 特徴抽出\nCPUで音高を抽出し(モデルに音高がある場合のみ)、GPUで特徴を抽出する(GPU番号を選択すべし)",
+ "Actually calculated": "実際計算",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "入力ソースの音量エンベロープと出力音量エンベロープの融合率 1に近づくほど、出力音量エンベロープの割合が高くなる",
+ "Algorithmic delays (ms)": "推論遅延(ms)",
+ "All processes have been completed!": "全工程が完了！",
+ "Audio device": "音声デバイス",
+ "Auto-detect index path and select from the dropdown": "索引パスの自動検出 ドロップダウンで選択",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "一括変換、変換する音声フォルダを入力、または複数の音声ファイルをアップロードし、指定したフォルダ(デフォルトのopt)に変換した音声を出力します。",
+ "Batch inference": "一括推論",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "UVR5モデルを使用したボーカル伴奏の分離バッチ処理。
有効なフォルダーパスフォーマットの例: D:\\path\\to\\input\\folder (エクスプローラーのアドレスバーからコピーします)。
モデルは三つのカテゴリに分かれています:
1. ボーカルを保持: ハーモニーのないオーディオに対してこれを選択します。HP5よりもボーカルをより良く保持します。HP2とHP3の二つの内蔵モデルが含まれています。HP3は伴奏をわずかに漏らす可能性がありますが、HP2よりもわずかにボーカルをより良く保持します。
2. 主なボーカルのみを保持: ハーモニーのあるオーディオに対してこれを選択します。主なボーカルを弱める可能性があります。HP5の一つの内蔵モデルが含まれています。
3. ディリバーブとディレイモデル (by FoxJoy):
(1) MDX-Net: ステレオリバーブの除去に最適な選択肢ですが、モノリバーブは除去できません；
(234) DeEcho: ディレイ効果を除去します。AggressiveモードはNormalモードよりも徹底的に除去します。DeReverbはさらにリバーブを除去し、モノリバーブを除去することができますが、高周波のリバーブが強い内容に対しては非常に効果的ではありません。
ディリバーブ/ディレイに関する注意点:
1. DeEcho-DeReverbモデルの処理時間は、他の二つのDeEchoモデルの約二倍です。
2. MDX-Net-Dereverbモデルは非常に遅いです。
3. 推奨される最もクリーンな設定は、最初にMDX-Netを適用し、その後にDeEcho-Aggressiveを適用することです。",
+ "Batch size per GPU": "GPUごとのバッチサイズ",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "すべての学習データをメモリにキャッシュするかどうか。10分以下の小さなデータはキャッシュして学習を高速化できますが、大きなデータをキャッシュするとメモリが破裂し、あまり速度が上がりません。",
+ "Calculate": "計算",
+ "Choose sample rate of the device": "デバイスサンプリング率を使用",
+ "Choose sample rate of the model": "モデルサンプリング率を使用",
+ "Convert": "変換",
+ "Device type": "デバイス種類",
+ "Enable phase vocoder": "启用相位声码器",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpeカード番号設定：異なるプロセスに使用するカード番号を入力する。例えば、0-0-1でカード0に2つのプロセス、カード1に1つのプロセスを実行する。",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "ハイフンで区切って使用するGPUの番号を入力します。例えば0-1-2はGPU0、GPU1、GPU2を使用します",
+ "Enter the experiment name": "モデル名",
+ "Enter the path of the audio folder to be processed": "処理するオーディオファイルのフォルダパスを入力してください",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "処理対象音声フォルダーのパスを入力してください(エクスプローラーのアドレスバーからコピーしてください)",
+ "Enter the path of the training folder": "学習用フォルダのパスを入力してください",
+ "Exist": "有",
+ "Export Onnx": "Onnx抽出",
+ "Export Onnx Model": "Onnxに変換",
+ "Export audio (click on the three dots in the lower right corner to download)": "出力音声(右下の三点をクリックしてダウンロードできます)",
+ "Export file format": "エクスポート形式",
+ "Extra inference time": "追加推論時間",
+ "Extract": "抽出",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0(最低共振周波数)カーブファイル(オプション、1行に1ピッチ、デフォルトのF0(最低共振周波数)とエレベーションを置き換えます。)",
+ "FAQ (Frequently Asked Questions)": "よくある質問",
+ "Fade length": "フェードイン/フェードアウト長",
+ "Fail": "失败",
+ "Feature extraction": "特徴抽出",
+ "Feature searching ratio": "検索特徴率",
+ "Formant offset": "共振偏移",
+ "Fusion": "マージ",
+ "GPU Information": "GPU情報",
+ "General settings": "一般設定",
+ "Hidden": "無表示",
+ "ID of model A (long)": "AモデルID(長)",
+ "ID of model B (long)": "BモデルID(長)",
+ "ID(long)": "ID(長)",
+ "ID(short)": "ID(短)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3 次に、harvestピッチの認識結果に対してメディアンフィルタを使用します。値はフィルター半径で、ミュートを減衰させるために使用します。",
+ "Inference time (ms)": "推論時間(ms)",
+ "Inferencing voice": "音源推論",
+ "Information": "情報",
+ "Input device": "入力デバイス",
+ "Input noise reduction": "入力騒音低減",
+ "Input voice monitor": "入力返聴",
+ "Link index to outside folder": "索引を外部フォルダへリンク",
+ "Load model": "モデルをロード",
+ "Load pre-trained base model D path": "事前学習済みのDモデルのパス",
+ "Load pre-trained base model G path": "事前学習済みのGモデルのパス",
+ "Loudness factor": "ラウドネス係数",
+ "Model": "モデル",
+ "Model Author": "モデル作者",
+ "Model Author (Nullable)": "モデル作者(空き可)",
+ "Model Inference": "モデル推論",
+ "Model architecture version": "モデルのバージョン",
+ "Model info": "モデル情報",
+ "Model information to be modified": "変更するモデル情報",
+ "Model information to be placed": "挿入するモデル情報",
+ "Model name": "モデル名",
+ "Modify": "変更",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "複数のオーディオファイルをインポートすることもできます。フォルダパスが存在する場合、この入力は無視されます。",
+ "No": "いいえ",
+ "None": "空き",
+ "Not exist": "無",
+ "Number of CPU processes used for harvest pitch algorithm": "harvestプロセス数",
+ "Number of CPU processes used for pitch extraction and data processing": "ピッチの抽出やデータ処理に使用するCPUスレッド数",
+ "One-click training": "ワンクリック学習",
+ "Onnx Export Path": "Onnx出力パス",
+ "Output converted voice": "出力音声変換",
+ "Output device": "出力デバイス",
+ "Output information": "出力情報",
+ "Output noise reduction": "出力騒音低減",
+ "Path to Model": "モデルパス",
+ "Path to Model A": "Aモデルのパス",
+ "Path to Model B": "Bモデルのパス",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "特徴検索ライブラリへのパス 空の場合はドロップダウンで選択",
+ "Performance settings": "パフォーマンス設定",
+ "Pitch detection algorithm": "ピッチアルゴリズム",
+ "Pitch guidance (f0)": "ピッチ導き(f0)",
+ "Pitch settings": "音程設定",
+ "Please choose the .index file": "indexファイルを選択してください",
+ "Please choose the .pth file": "pthファイルを選択してください",
+ "Please specify the speaker/singer ID": "話者IDを指定してください",
+ "Process data": "データ処理",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "明確な子音と呼吸音を保護し、電子音の途切れやその他のアーティファクトを防止します。0.5でオフになります。下げると保護が強化されますが、indexの効果が低下する可能性があります。",
+ "RVC Model Path": "RVCモデルパス",
+ "Read from model": "モデルから読込",
+ "Refresh voice list and index path": "音源リストと索引パスの更新",
+ "Reload device list": "デバイス列の再読込",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "最終的なサンプリング率へのポストプロセッシングのリサンプリング リサンプリングしない場合は0",
+ "Response threshold": "反応閾値",
+ "Sample length": "サンプル長",
+ "Sampling rate": "サンプル率",
+ "Save a small final model to the 'weights' folder at each save point": "各保存時点の小モデルを全部weightsフォルダに保存するかどうか",
+ "Save file name (default: same as the source file)": "保存するファイル名、デフォルトでは空欄で元のファイル名と同じ名前になります",
+ "Save frequency (save_every_epoch)": "エポックごとの保存頻度",
+ "Save name": "保存ファイル名",
+ "Save only the latest '.ckpt' file to save disk space": "ハードディスク容量を節約するため、最新のckptファイルのみを保存しますか？",
+ "Saved model name (without extension)": "拡張子のない保存するモデル名",
+ "Sealing date": "締め日付",
+ "Select Speaker/Singer ID": "話者IDを選択してください",
+ "Select the .index file": ".indexファイルを選択",
+ "Select the .pth file": ".pthファイルを選択",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "ピッチ抽出アルゴリズムの選択、歌声はpmで高速化でき、harvestは低音が良いが信じられないほど遅く、crepeは良く動くがGPUを喰います",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "ピッチ抽出アルゴリズムの選択：歌声はpmで高速化でき、入力した音声が高音質でCPUが貧弱な場合はdioで高速化でき、harvestの方が良いが遅く、rmvpeがベストだがCPU/GPUを若干食います。",
+ "Similarity": "相似度",
+ "Similarity (from 0 to 1)": "相似度(0～1)",
+ "Single inference": "一度推論",
+ "Specify output folder": "出力フォルダを指定してください",
+ "Specify the output folder for accompaniment": "マスター以外の出力音声フォルダーを指定する",
+ "Specify the output folder for vocals": "マスターの出力音声フォルダーを指定する",
+ "Start audio conversion": "音声変換を開始",
+ "Step 1: Processing data": "step1:処理中のデータ",
+ "Step 3a: Model training started": "step3a:学習中のモデル",
+ "Stop audio conversion": "音声変換を停止",
+ "Successfully built index into": "成功构建索引到",
+ "Takeover WASAPI device": "WASAPIデバイスを独占",
+ "Target sample rate": "目標サンプリング率",
+ "The audio file to be processed": "処理待ち音声",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "本ソフトウェアはMITライセンスに基づくオープンソースであり、製作者は本ソフトウェアに対していかなる責任を持ちません。本ソフトウェアの利用者および本ソフトウェアから派生した音源(成果物)を配布する者は、本ソフトウェアに対して自身で責任を負うものとします。
この条項に同意しない場合、パッケージ内のコードやファイルを使用や参照を禁じます。詳しくはLICENSEをご覧ください。",
+ "Total training epochs (total_epoch)": "総エポック数",
+ "Train": "学習",
+ "Train feature index": "特徴索引の学習",
+ "Train model": "モデルの学習",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "学習終了時に、学習ログやフォルダ内のtrain.logを確認することができます",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "ピッチ変更(整数、半音数、上下オクターブ12-12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "学習に対応したGPUが動作しないのは残念です。",
+ "Unknown": "未知",
+ "Unload model to save GPU memory": "音源を削除してメモリを節約",
+ "Version": "バージョン",
+ "View": "表示",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "伴奏ボーカル分離＆残響除去＆エコー除去",
+ "Weight (w) for Model A": "Aモデル占有率",
+ "Whether the model has pitch guidance": "モデルに音高ガイドを付けるかどうか",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "モデルに音高ガイドを付けるかどうか、1は付ける、0は付けない",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "モデルに音高ガイドがあるかどうか(歌唱には必要ですが、音声には必要ありません)",
+ "Yes": "はい",
+ "ckpt Processing": "ckptファイルの処理",
+ "index path cannot contain unicode characters": "indexファイルのパスに漢字を含んではいけません",
+ "pth path cannot contain unicode characters": "pthファイルのパスに漢字を含んではいけません",
+ "step2:Pitch extraction & feature extraction": "step2:ピッチ抽出と特徴抽出"
+}
diff --git a/i18n/locale/ko_KR.json b/i18n/locale/ko_KR.json
new file mode 100644
index 0000000000000000000000000000000000000000..73398f37f971067f52d48d21879bbdd5e4caa8af
--- /dev/null
+++ b/i18n/locale/ko_KR.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### 모델 추출\n> logs 폴더 아래의 큰 파일 모델 경로 입력\n\n훈련 중간에 중단한 모델의 자동 추출 및 소형 파일 모델 저장이 안 되거나 중간 모델을 테스트하고 싶은 경우에 적합",
+ "### Model fusion\nCan be used to test timbre fusion.": "### 모델 융합\n음색 융합 테스트에 사용 가능",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 모델 정보 수정\n> 오직 weights 폴더 아래에서 추출된 작은 모델 파일만 지원",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. 실험 구성 작성\n실험 데이터는 logs에 저장, 각 실험은 하나의 폴더, 수동으로 실험 이름 경로 입력 필요, 실험 구성, 로그, 훈련된 모델 파일 포함.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 모델 정보 보기\n> 오직 weights 폴더에서 추출된 소형 모델 파일만 지원",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "입력 소스 볼륨 엔벨로프와 출력 볼륨 엔벨로프의 결합 비율 입력, 1에 가까울수록 출력 엔벨로프 사용",
+ "Algorithmic delays (ms)": "알고리즘 지연(ms)",
+ "All processes have been completed!": "전체 과정 완료!",
+ "Audio device": "오디오 장치",
+ "Auto-detect index path and select from the dropdown": "자동으로 index 경로 감지, 드롭다운 선택(dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "일괄 변환, 변환할 오디오 파일 폴더 입력 또는 여러 오디오 파일 업로드, 지정된 폴더(기본값 opt)에 변환된 오디오 출력.",
+ "Batch inference": "일괄 추론",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "인간 목소리와 반주 분리 배치 처리, UVR5 모델 사용.
적절한 폴더 경로 예시: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(파일 관리자 주소 표시줄에서 복사하면 됨).
모델은 세 가지 유형으로 나뉨:
1. 인간 목소리 보존: 화음이 없는 오디오에 이것을 선택, HP5보다 주된 인간 목소리 보존에 더 좋음. 내장된 HP2와 HP3 두 모델, HP3는 약간의 반주 누락 가능성이 있지만 HP2보다 주된 인간 목소리 보존이 약간 더 좋음;
2. 주된 인간 목소리만 보존: 화음이 있는 오디오에 이것을 선택, 주된 인간 목소리에 약간의 약화 가능성 있음. 내장된 HP5 모델 하나;
3. 혼효음 제거, 지연 제거 모델(by FoxJoy):
(1)MDX-Net(onnx_dereverb): 이중 채널 혼효음에는 최선의 선택, 단일 채널 혼효음은 제거할 수 없음;
(234)DeEcho: 지연 제거 효과. Aggressive는 Normal보다 더 철저하게 제거, DeReverb는 추가로 혼효음을 제거, 단일 채널 혼효음은 제거 가능하지만 고주파 중심의 판 혼효음은 완전히 제거하기 어려움.
혼효음/지연 제거, 부록:
1. DeEcho-DeReverb 모델의 처리 시간은 다른 두 개의 DeEcho 모델의 거의 2배임;
2. MDX-Net-Dereverb 모델은 상당히 느림;
3. 개인적으로 추천하는 가장 깨끗한 구성은 MDX-Net 다음에 DeEcho-Aggressive 사용.",
+ "Batch size per GPU": "각 그래픽 카드의 batch_size",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "모든 훈련 세트를 VRAM에 캐시할지 여부. 10분 미만의 소량 데이터는 캐시하여 훈련 속도를 높일 수 있지만, 대량 데이터 캐시는 VRAM을 과부하시키고 속도를 크게 향상시키지 못함",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "장치 샘플링 레이트 사용",
+ "Choose sample rate of the model": "모델 샘플링 레이트 사용",
+ "Convert": "변환",
+ "Device type": "设备类型",
+ "Enable phase vocoder": "위상 보코더 활성화",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpe 카드 번호 설정: -로 구분된 입력 사용 카드 번호, 예: 0-0-1은 카드 0에서 2개 프로세스, 카드 1에서 1개 프로세스 실행",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "-로 구분하여 입력하는 카드 번호, 예: 0-1-2는 카드 0, 카드 1, 카드 2 사용",
+ "Enter the experiment name": "실험명 입력",
+ "Enter the path of the audio folder to be processed": "처리할 오디오 파일 폴더 경로 입력",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "처리할 오디오 파일 폴더 경로 입력(파일 탐색기 주소 표시줄에서 복사)",
+ "Enter the path of the training folder": "훈련 파일 폴더 경로 입력",
+ "Exist": "有",
+ "Export Onnx": "Onnx 내보내기",
+ "Export Onnx Model": "Onnx 모델 내보내기",
+ "Export audio (click on the three dots in the lower right corner to download)": "출력 오디오(오른쪽 하단 세 개의 점, 클릭하면 다운로드 가능)",
+ "Export file format": "내보낼 파일 형식",
+ "Extra inference time": "추가 추론 시간",
+ "Extract": "추출",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0 곡선 파일, 선택적, 한 줄에 하나의 피치, 기본 F0 및 음높이 조절 대체",
+ "FAQ (Frequently Asked Questions)": "자주 묻는 질문",
+ "Fade length": "페이드 인/아웃 길이",
+ "Fail": "失败",
+ "Feature extraction": "특징 추출",
+ "Feature searching ratio": "검색 특징 비율",
+ "Formant offset": "共振偏移",
+ "Fusion": "융합",
+ "GPU Information": "그래픽 카드 정보",
+ "General settings": "일반 설정",
+ "Hidden": "Hidden",
+ "ID of model A (long)": "ID of model A (long)",
+ "ID of model B (long)": "ID of model B (long)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3인 경우 harvest 피치 인식 결과에 중간값 필터 적용, 필터 반경은 값으로 지정, 사용 시 무성음 감소 가능",
+ "Inference time (ms)": "추론 시간 (ms)",
+ "Inferencing voice": "추론 음색",
+ "Information": "정보",
+ "Input device": "입력 장치",
+ "Input noise reduction": "입력 노이즈 감소",
+ "Input voice monitor": "입력 모니터링",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "모델 로드",
+ "Load pre-trained base model D path": "미리 훈련된 베이스 모델 D 경로 로드",
+ "Load pre-trained base model G path": "미리 훈련된 베이스 모델 G 경로 로드",
+ "Loudness factor": "음량 인자",
+ "Model": "모델",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "모델 추론",
+ "Model architecture version": "모델 버전 및 모델",
+ "Model info": "모델 정보",
+ "Model information to be modified": "변경할 모델 정보",
+ "Model information to be placed": "삽입할 모델 정보",
+ "Model name": "Model name",
+ "Modify": "수정",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "여러 오디오 파일을 일괄 입력할 수도 있음, 둘 중 하나 선택, 폴더 우선 읽기",
+ "No": "아니오",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "harvest 프로세스 수",
+ "Number of CPU processes used for pitch extraction and data processing": "음높이 추출 및 데이터 처리에 사용되는 CPU 프로세스 수",
+ "One-click training": "원클릭 훈련",
+ "Onnx Export Path": "Onnx 출력 경로",
+ "Output converted voice": "출력 음성 변조",
+ "Output device": "출력 장치",
+ "Output information": "출력 정보",
+ "Output noise reduction": "출력 노이즈 감소",
+ "Path to Model": "모델 경로",
+ "Path to Model A": "A 모델 경로",
+ "Path to Model B": "B 모델 경로",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "특징 검색 라이브러리 파일 경로, 비어 있으면 드롭다운 선택 결과 사용",
+ "Performance settings": "성능 설정",
+ "Pitch detection algorithm": "음높이 알고리즘",
+ "Pitch guidance (f0)": "音高引导(f0)",
+ "Pitch settings": "음조 설정",
+ "Please choose the .index file": "index 파일 선택",
+ "Please choose the .pth file": "pth 파일 선택",
+ "Please specify the speaker/singer ID": "화자 ID 지정 필요",
+ "Process data": "데이터 처리",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "청자음과 호흡 소리를 보호, 전자음 찢김 등의 아티팩트 방지, 0.5까지 올려서 비활성화, 낮추면 보호 강도 증가하지만 인덱스 효과 감소 가능성 있음",
+ "RVC Model Path": "RVC 모델 경로",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "음색 목록 및 인덱스 경로 새로고침",
+ "Reload device list": "장치 목록 재로드",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "후처리 재샘플링을 최종 샘플링 레이트로, 0은 재샘플링하지 않음",
+ "Response threshold": "응답 임계값",
+ "Sample length": "샘플링 길이",
+ "Sampling rate": "샘플링률",
+ "Save a small final model to the 'weights' folder at each save point": "저장 시마다 최종 소형 모델을 weights 폴더에 저장할지 여부",
+ "Save file name (default: same as the source file)": "저장될 파일명, 기본적으로 빈 공간은 원본 파일과 동일한 이름으로",
+ "Save frequency (save_every_epoch)": "저장 빈도 save_every_epoch",
+ "Save name": "저장 이름",
+ "Save only the latest '.ckpt' file to save disk space": "디스크 공간을 절약하기 위해 최신 ckpt 파일만 저장할지 여부",
+ "Saved model name (without extension)": "저장된 모델명은 접미사 없음",
+ "Sealing date": "Sealing date",
+ "Select Speaker/Singer ID": "화자 ID 선택",
+ "Select the .index file": ".index 파일 선택",
+ "Select the .pth file": ".pth 파일 선택",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "음높이 추출 알고리즘 선택, 노래 입력 시 pm으로 속도 향상, harvest는 저음이 좋지만 매우 느림, crepe는 효과가 좋지만 GPU 사용, rmvpe는 효과가 가장 좋으며 GPU를 적게 사용",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "음높이 추출 알고리즘 선택: 노래 입력 시 pm으로 속도 향상, 고품질 음성에는 CPU가 부족할 때 dio 사용, harvest는 품질이 더 좋지만 느림, rmvpe는 효과가 가장 좋으며 CPU/GPU를 적게 사용",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "단일 추론",
+ "Specify output folder": "출력 파일 폴더 지정",
+ "Specify the output folder for accompaniment": "주된 목소리가 아닌 출력 폴더 지정",
+ "Specify the output folder for vocals": "주된 목소리 출력 폴더 지정",
+ "Start audio conversion": "오디오 변환 시작",
+ "Step 1: Processing data": "step1: 데이터 처리 중",
+ "Step 3a: Model training started": "step3a: 모델 훈련 중",
+ "Stop audio conversion": "오디오 변환 중지",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "목표 샘플링률",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "이 소프트웨어는 MIT 라이선스로 공개되며, 저자는 소프트웨어에 대해 어떠한 통제권도 가지지 않습니다. 모든 귀책사유는 소프트웨어 사용자 및 소프트웨어에서 생성된 결과물을 사용하는 당사자에게 있습니다.
해당 조항을 인정하지 않는 경우, 소프트웨어 패키지의 어떠한 코드나 파일도 사용하거나 인용할 수 없습니다. 자세한 내용은 루트 디렉토리의 LICENSE를 참조하세요.",
+ "Total training epochs (total_epoch)": "총 훈련 라운드 수 total_epoch",
+ "Train": "훈련",
+ "Train feature index": "특징 인덱스 훈련",
+ "Train model": "모델 훈련",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "훈련 완료, 콘솔 훈련 로그 또는 실험 폴더 내의 train.log 확인 가능",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "키 변경(정수, 반음 수, 옥타브 상승 12, 옥타브 하강 -12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "사용 가능한 그래픽 카드가 없어 훈련을 지원할 수 없습니다",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "음색 언로드로 디스플레이 메모리 절약",
+ "Version": "버전",
+ "View": "보기",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "반주 인간 목소리 분리 & 혼효음 제거 & 에코 제거",
+ "Weight (w) for Model A": "A 모델 가중치",
+ "Whether the model has pitch guidance": "모델이 음높이 지도를 포함하는지 여부",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "모델이 음높이 지도를 포함하는지 여부, 1은 예, 0은 아니오",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "모델이 음높이 지도를 포함하는지 여부(노래에는 반드시 필요, 음성에는 필요 없음)",
+ "Yes": "예",
+ "ckpt Processing": "ckpt 처리",
+ "index path cannot contain unicode characters": "index 파일 경로는 중국어를 포함할 수 없음",
+ "pth path cannot contain unicode characters": "pth 파일 경로는 중국어를 포함할 수 없음",
+ "step2:Pitch extraction & feature extraction": "step2: 음높이 추출 & 특징 추출 중"
+}
diff --git a/i18n/locale/pt_BR.json b/i18n/locale/pt_BR.json
new file mode 100644
index 0000000000000000000000000000000000000000..1fa9560c24638c4850ebfa2e94a5cf858cb34c59
--- /dev/null
+++ b/i18n/locale/pt_BR.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Extração do modelo\n> Insira o caminho do modelo de arquivo grande na pasta 'logs'.\n\nIsso é útil se você quiser interromper o treinamento no meio do caminho e extrair e salvar manualmente um arquivo de modelo pequeno, ou se quiser testar um modelo intermediário.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Modificar informações do modelo\n> Suportado apenas para arquivos de modelo pequenos extraídos da pasta 'weights'.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Etapa 1. Preencha a configuração experimental.\nOs dados experimentais são armazenados na pasta 'logs', com cada experimento tendo uma pasta separada. Digite manualmente o caminho do nome do experimento, que contém a configuração experimental, os logs e os arquivos de modelo treinados.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Exibir informações do modelo\n> Suportado apenas para arquivos de modelo pequenos extraídos da pasta 'weights'.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "O envelope de volume da fonte de entrada substitui a taxa de fusão do envelope de volume de saída, quanto mais próximo de 1, mais o envelope de saída é usado:",
+ "Algorithmic delays (ms)": "Atrasos algorítmicos (ms)",
+ "All processes have been completed!": "Todos os processos foram concluídos!",
+ "Audio device": "Audio device",
+ "Auto-detect index path and select from the dropdown": "Detecte automaticamente o caminho do Index e selecione no menu suspenso:",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Conversão em Massa.",
+ "Batch inference": "Conversão em Lote",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Processamento em lote para separação de acompanhamento vocal usando o modelo UVR5.
Exemplo de um formato de caminho de pasta válido: D:\\caminho\\para a pasta\\entrada\\ (copie-o da barra de endereços do gerenciador de arquivos).
O modelo é dividido em três categorias:
1. Preservar vocais: Escolha esta opção para áudio sem harmonias. Ele preserva os vocais melhor do que o HP5. Inclui dois modelos integrados: HP2 e HP3. O HP3 pode vazar ligeiramente o acompanhamento, mas preserva os vocais um pouco melhor do que o HP2.
2 Preservar apenas os vocais principais: Escolha esta opção para áudio com harmonias. Isso pode enfraquecer os vocais principais. Ele inclui um modelo embutido: HP5.
3. Modelos de de-reverb e de-delay (por FoxJoy):
(1) MDX-Net: A melhor escolha para remoção de reverb estéreo, mas não pode remover reverb mono;
(234) DeEcho: Remove efeitos de atraso. O modo agressivo remove mais completamente do que o modo normal. O DeReverb também remove reverb e pode remover reverb mono, mas não de forma muito eficaz para conteúdo de alta frequência fortemente reverberado.
Notas de de-reverb/de-delay:
1. O tempo de processamento para o modelo DeEcho-DeReverb é aproximadamente duas vezes maior que os outros dois modelos DeEcho.
2 O modelo MDX-Net-Dereverb é bastante lento.
3. A configuração mais limpa recomendada é aplicar MDX-Net primeiro e depois DeEcho-Aggressive.",
+ "Batch size per GPU": "Batch Size (DEIXE COMO ESTÁ a menos que saiba o que está fazendo, no Colab pode deixar até 20!):",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Se deve armazenar em cache todos os conjuntos de treinamento na memória de vídeo. Pequenos dados com menos de 10 minutos podem ser armazenados em cache para acelerar o treinamento, e um cache de dados grande irá explodir a memória de vídeo e não aumentar muito a velocidade:",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "Choose sample rate of the device",
+ "Choose sample rate of the model": "Choose sample rate of the model",
+ "Convert": "Converter",
+ "Device type": "Device type",
+ "Enable phase vocoder": "Enable phase vocoder",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Configuração do número do cartão rmvpe: Use - para separar os números dos cartões de entrada de diferentes processos. Por exemplo, 0-0-1 é usado para executar 2 processos no cartão 0 e 1 processo no cartão 1.",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Digite o (s) índice(s) da GPU separados por '-', por exemplo, 0-1-2 para usar a GPU 0, 1 e 2:",
+ "Enter the experiment name": "Nome da voz:",
+ "Enter the path of the audio folder to be processed": "Caminho da pasta de áudio a ser processada:",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Caminho da pasta de áudio a ser processada (copie-o da barra de endereços do gerenciador de arquivos):",
+ "Enter the path of the training folder": "Caminho da pasta de treinamento:",
+ "Exist": "Exist",
+ "Export Onnx": "Exportar Onnx",
+ "Export Onnx Model": "Exportar Modelo Onnx",
+ "Export audio (click on the three dots in the lower right corner to download)": "Exportar áudio (clique nos três pontos no canto inferior direito para baixar)",
+ "Export file format": "Qual formato de arquivo você prefere?",
+ "Extra inference time": "Tempo extra de inferência",
+ "Extract": "Extrato",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "Arquivo de curva F0 (opcional). Um arremesso por linha. Substitui a modulação padrão F0 e tom:",
+ "FAQ (Frequently Asked Questions)": "FAQ (Perguntas frequentes)",
+ "Fade length": "Comprimento de desvanecimento",
+ "Fail": "失败",
+ "Feature extraction": "Extrair Tom",
+ "Feature searching ratio": "Taxa de recurso de recuperação:",
+ "Formant offset": "共振偏移",
+ "Fusion": "Fusão",
+ "GPU Information": "Informações da GPU",
+ "General settings": "Configurações gerais",
+ "Hidden": "不显示",
+ "ID of model A (long)": "A模型ID(长)",
+ "ID of model B (long)": "B模型ID(长)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3, use o filtro mediano para o resultado do reconhecimento do tom da heverst, e o valor é o raio do filtro, que pode enfraquecer o mudo.",
+ "Inference time (ms)": "Tempo de inferência (ms)",
+ "Inferencing voice": "Escolha o seu Modelo:",
+ "Information": "信息",
+ "Input device": "Dispositivo de entrada",
+ "Input noise reduction": "Redução de ruído de entrada",
+ "Input voice monitor": "Monitoramento de entrada",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "Modelo",
+ "Load pre-trained base model D path": "Carregue o caminho D do modelo base pré-treinado:",
+ "Load pre-trained base model G path": "Carregue o caminho G do modelo base pré-treinado:",
+ "Loudness factor": "Fator de volume",
+ "Model": "Modelo",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "Inference",
+ "Model architecture version": "Versão:",
+ "Model info": "Model info",
+ "Model information to be modified": "Informações do modelo a ser modificado:",
+ "Model information to be placed": "Informações do modelo a ser colocado:",
+ "Model name": "Model name",
+ "Modify": "Editar",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Você também pode inserir arquivos de áudio em lotes. Escolha uma das duas opções. É dada prioridade à leitura da pasta.",
+ "No": "Não",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "Número de processos harvest",
+ "Number of CPU processes used for pitch extraction and data processing": "Número de processos de CPU usados para extração de tom e processamento de dados:",
+ "One-click training": "Treinamento com um clique",
+ "Onnx Export Path": "Caminho de exportação ONNX:",
+ "Output converted voice": "Mudança de voz de saída",
+ "Output device": "Dispositivo de saída",
+ "Output information": "Informação de saída",
+ "Output noise reduction": "Redução de ruído de saída",
+ "Path to Model": "Caminho para o Modelo:",
+ "Path to Model A": "Caminho para o Modelo A:",
+ "Path to Model B": "Caminho para o Modelo B:",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Caminho para o arquivo de Index. Deixe em branco para usar o resultado selecionado no menu debaixo:",
+ "Performance settings": "Configurações de desempenho.",
+ "Pitch detection algorithm": "Algoritmo de detecção de pitch",
+ "Pitch guidance (f0)": "Pitch guidance (f0)",
+ "Pitch settings": "Configurações de tom",
+ "Please choose the .index file": "Selecione o arquivo de Index",
+ "Please choose the .pth file": "Selecione o arquivo pth",
+ "Please specify the speaker/singer ID": "Especifique o ID do locutor/cantor:",
+ "Process data": "Processar o Conjunto de Dados",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Proteja consoantes sem voz e sons respiratórios, evite artefatos como quebra de som eletrônico e desligue-o quando estiver cheio de 0,5. Diminua-o para aumentar a proteção, mas pode reduzir o efeito de indexação:",
+ "RVC Model Path": "Caminho do Modelo RVC:",
+ "Read from model": "Read from model",
+ "Refresh voice list and index path": "Atualizar lista de voz e caminho do Index",
+ "Reload device list": "Recarregar lista de dispositivos",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Reamostragem pós-processamento para a taxa de amostragem final, 0 significa sem reamostragem:",
+ "Response threshold": "Limiar de resposta",
+ "Sample length": "Comprimento da Amostra",
+ "Sampling rate": "Sampling rate",
+ "Save a small final model to the 'weights' folder at each save point": "Salve um pequeno modelo final na pasta 'weights' em cada ponto de salvamento:",
+ "Save file name (default: same as the source file)": "Salvar nome do arquivo (padrão: igual ao arquivo de origem):",
+ "Save frequency (save_every_epoch)": "Faça backup a cada # de Epoch:",
+ "Save name": "Salvar nome",
+ "Save only the latest '.ckpt' file to save disk space": "Só deve salvar apenas o arquivo ckpt mais recente para economizar espaço em disco:",
+ "Saved model name (without extension)": "Nome do modelo salvo (sem extensão):",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "Selecione Palestrantes/Cantores ID:",
+ "Select the .index file": "Selecione o Index",
+ "Select the .pth file": "Selecione o Arquivo",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Selecione o algoritmo de extração de tom \n'pm': extração mais rápida, mas discurso de qualidade inferior; \n'harvest': graves melhores, mas extremamente lentos; \n'harvest': melhor qualidade, mas extração mais lenta); 'crepe': melhor qualidade, mas intensivo em GPU; 'magio-crepe': melhor opção; 'RMVPE': um modelo robusto para estimativa de afinação vocal em música polifônica;",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Selecione o algoritmo de extração de tom \n'pm': extração mais rápida, mas discurso de qualidade inferior; \n'harvest': graves melhores, mas extremamente lentos; \n'crepe': melhor qualidade (mas intensivo em GPU);\n rmvpe tem o melhor efeito e consome menos CPU/GPU.",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "Único",
+ "Specify output folder": "Especifique a pasta de saída:",
+ "Specify the output folder for accompaniment": "Informar a pasta de saída para acompanhamento:",
+ "Specify the output folder for vocals": "Especifique a pasta de saída para vocais:",
+ "Start audio conversion": "Iniciar conversão de áudio",
+ "Step 1: Processing data": "Etapa 1: Processamento de dados",
+ "Step 3a: Model training started": "Etapa 3a: Treinamento do modelo iniciado",
+ "Stop audio conversion": "Conversão de áudio",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "Taxa de amostragem:",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "The Mangio-RVC 💻 | Tradução por Krisp e Rafael Godoy Ebert | AI HUB BRASIL
Este software é de código aberto sob a licença MIT. O autor não tem qualquer controle sobre o software. Aqueles que usam o software e divulgam os sons exportados pelo software são totalmente responsáveis.
Se você não concorda com este termo, você não pode usar ou citar nenhum código e arquivo no pacote de software. Para obter detalhes, consulte o diretório raiz O acordo a ser seguido para uso LICENSE",
+ "Total training epochs (total_epoch)": "Número total de ciclos(epoch) de treino (se escolher um valor alto demais, o seu modelo parecerá terrivelmente sobretreinado):",
+ "Train": "Treinar",
+ "Train feature index": "Treinar Index",
+ "Train model": "Treinar Modelo",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Após o término do treinamento, você pode verificar o log de treinamento do console ou train.log na pasta de experimentos",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Mude o tom aqui. Se a voz for do mesmo sexo, não é necessario alterar (12 caso seja Masculino para feminino, -12 caso seja ao contrário).",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Infelizmente, não há GPU compatível disponível para apoiar o seu treinamento.",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "Descarregue a voz para liberar a memória da GPU:",
+ "Version": "Versão",
+ "View": "Visualizar",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "UVR5",
+ "Weight (w) for Model A": "Peso (w) para o modelo A:",
+ "Whether the model has pitch guidance": "Se o modelo tem orientação de tom:",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Se o modelo tem orientação de passo (1: sim, 0: não):",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Se o modelo tem orientação de tom (necessário para cantar, opcional para fala):",
+ "Yes": "Sim",
+ "ckpt Processing": "processamento ckpt",
+ "index path cannot contain unicode characters": "O caminho do arquivo de Index não pode conter caracteres chineses",
+ "pth path cannot contain unicode characters": "o caminho do arquivo pth não pode conter caracteres chineses",
+ "step2:Pitch extraction & feature extraction": "step2:Pitch extraction & feature extraction"
+}
diff --git a/i18n/locale/ru_RU.json b/i18n/locale/ru_RU.json
new file mode 100644
index 0000000000000000000000000000000000000000..263b82d5898bfb9fb1e2ed875737fd7ae06c7858
--- /dev/null
+++ b/i18n/locale/ru_RU.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Создание модели из данных\n> Полученных в процессе обучения (введите путь к большому файлу модели в папке 'logs').\n\nМожет пригодиться, если вам нужно завершить обучение и получить маленький файл готовой модели, или если вам нужно проверить недообученную модель.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Слияние моделей> Может быть использовано для проверки слияния тембра.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Изменить информацию о модели\n> Работает только с маленькими моделями, взятыми из папки 'weights'.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Шаг 1. Конфигурирование модели.\nДанные обучения модели сохраняются в папку 'logs', и для каждой модели создаётся отдельная папка. Введите вручную путь к настройкам для модели, в которой находятся логи и тренировочные файлы.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Просмотреть информацию о модели\n> Работает только с маленькими моделями, взятыми из папки 'weights'.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Использовать громкость входного файла для замены или перемешивания с громкостью выходного файла. Чем ближе соотношение к 1, тем больше используется звука из выходного файла:",
+ "Algorithmic delays (ms)": "Algorithmic delays (ms)",
+ "All processes have been completed!": "Все процессы завершены!",
+ "Audio device": "Аудиоустройство",
+ "Auto-detect index path and select from the dropdown": "Автоматически найденные файлы индексов черт (выберите вариант из списка):",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Массовое преобразование. Введите путь к папке, в которой находятся файлы для преобразования голоса или выгрузите несколько аудиофайлов. Сконвертированные файлы будут сохранены в указанной папке (по умолчанию: 'opt').",
+ "Batch inference": "Batch inference",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Пакетная обработка для разделения вокального сопровождения с использованием модели UVR5.
Пример допустимого формата пути к папке: D:\\path\\to\\input\\folder
Модель разделена на три категории:
1. Сохранить вокал: выберите этот вариант для звука без гармоний. Он сохраняет вокал лучше, чем HP5. Он включает в себя две встроенные модели: HP2 и HP3. HP3 может немного пропускать инструментал, но сохраняет вокал немного лучше, чем HP2.
2. Сохранить только основной вокал: выберите этот вариант для звука с гармониями. Это может ослабить основной вокал. Он включает одну встроенную модель: HP5.
3. Модели удаления реверберации и задержки (от FoxJoy):
(1) MDX-Net: лучший выбор для удаления стереореверберации, но он не может удалить монореверберацию;
(234) DeEcho: удаляет эффекты задержки. Агрессивный режим удаляет более тщательно, чем Нормальный режим. DeReverb дополнительно удаляет реверберацию и может удалять монореверберацию, но не очень эффективно для сильно реверберированного высокочастотного контента.
Примечания по удалению реверберации/задержки:
1. Время обработки для модели DeEcho-DeReverb примерно в два раза больше, чем для двух других моделей DeEcho.
2. Модель MDX-Net-Dereverb довольно медленная.
3. Рекомендуемая самая чистая конфигурация — сначала применить MDX-Net, а затем DeEcho-Aggressive.",
+ "Batch size per GPU": "Размер пачки для GPU:",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Кэшировать все тренировочные сеты в видеопамять. Кэширование маленький датасетов (меньше 10 минут) может ускорить тренировку, но кэширование больших, наоборот, займёт много видеопамяти и не сильно ускорит тренировку:",
+ "Calculate": "Calculate",
+ "Choose sample rate of the device": "Choose sample rate of the device",
+ "Choose sample rate of the model": "Choose sample rate of the model",
+ "Convert": "Преобразовать",
+ "Device type": "Device type",
+ "Enable phase vocoder": "Enable phase vocoder",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Введите номера графических процессоров, разделенные символом «-», например, 0-0-1, чтобы запустить два процесса на GPU 0 и один процесс на GPU 1:",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "Введите, какие(-ую) GPU(-у) хотите использовать через '-', например 0-1-2, чтобы использовать GPU с номерами 0, 1 и 2:",
+ "Enter the experiment name": "Название модели:",
+ "Enter the path of the audio folder to be processed": "Путь к папке с аудиофайлами для обработки:",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "Путь к папке с аудиофайлами для переработки (можно скопировать путь из адресной строки файлового менеджера):",
+ "Enter the path of the training folder": "Путь к папке с аудиозаписями, на которых будет обучаться модель:",
+ "Exist": "Exist",
+ "Export Onnx": "Экспорт ONNX",
+ "Export Onnx Model": "Экспортировать модель",
+ "Export audio (click on the three dots in the lower right corner to download)": "Аудиофайл (чтобы скачать, нажмите на три точки справа в плеере)",
+ "Export file format": "Формат выходных файлов",
+ "Extra inference time": "Доп. время переработки",
+ "Extract": "Создать модель",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "Файл дуги F0 (не обязательно). Одна тональность на каждую строчку. Заменяет обычный F0 и модуляцию тональности:",
+ "FAQ (Frequently Asked Questions)": "ЧаВо (часто задаваемые вопросы)",
+ "Fade length": "Длина затухания",
+ "Fail": "失败",
+ "Feature extraction": "Извлечь черты",
+ "Feature searching ratio": "Соотношение поиска черт:",
+ "Formant offset": "共振偏移",
+ "Fusion": "Запустить слияние",
+ "GPU Information": "Информация о графических процессорах (GPUs):",
+ "General settings": "Основные настройки",
+ "Hidden": "Hidden",
+ "ID of model A (long)": "ID of model A (long)",
+ "ID of model B (long)": "ID of model B (long)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "Если значение больше 3: применить медианную фильтрацию к вытащенным тональностям. Значение контролирует радиус фильтра и может уменьшить излишнее дыхание.",
+ "Inference time (ms)": "Время переработки (мс)",
+ "Inferencing voice": "Желаемый голос:",
+ "Information": "Information",
+ "Input device": "Входное устройство",
+ "Input noise reduction": "Уменьшение входного шума",
+ "Input voice monitor": "Input voice monitor",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "Загрузить модель",
+ "Load pre-trained base model D path": "Путь к предварительно обученной базовой модели D:",
+ "Load pre-trained base model G path": "Путь к предварительно обученной базовой модели G:",
+ "Loudness factor": "коэффициент громкости",
+ "Model": "Модели",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "Изменение голоса",
+ "Model architecture version": "Версия архитектуры модели:",
+ "Model info": "Model info",
+ "Model information to be modified": "Информация, которая будет изменена:",
+ "Model information to be placed": "Информация о модели:",
+ "Model name": "Model name",
+ "Modify": "Изменить",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Можно также импортировать несколько аудиофайлов. Если путь к папке существует, то этот ввод игнорируется.",
+ "No": "Нет",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "Количество процессор harvest",
+ "Number of CPU processes used for pitch extraction and data processing": "Число процессов ЦП, используемое для оценки высоты голоса и обработки данных:",
+ "One-click training": "Обучение в одно нажатие",
+ "Onnx Export Path": "Путь для сохранения модели в формате ONNX:",
+ "Output converted voice": "Output converted voice",
+ "Output device": "Выходное устройство",
+ "Output information": "Статистика",
+ "Output noise reduction": "Уменьшение выходного шума",
+ "Path to Model": "Путь к папке:",
+ "Path to Model A": "Путь к модели А:",
+ "Path to Model B": "Путь к модели Б:",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Путь к файлу индекса черт. Оставьте пустым, чтобы использовать выбранный вариант из списка ниже:",
+ "Performance settings": "Настройки быстроты",
+ "Pitch detection algorithm": "Алгоритм оценки высоты звука",
+ "Pitch guidance (f0)": "Pitch guidance (f0)",
+ "Pitch settings": "Настройка высоты звука",
+ "Please choose the .index file": "Пожалуйста, выберите файл индекса",
+ "Please choose the .pth file": "Пожалуйста, выберите файл pth",
+ "Please specify the speaker/singer ID": "Номер говорящего/поющего:",
+ "Process data": "Обработать данные",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Защитить глухие согласные и звуки дыхания для предотвращения артефактов, например, разрывания в электронной музыке. Поставьте на 0.5, чтобы выключить. Уменьшите значение для повышения защиты, но учтите, что при этом может ухудшиться точность индексирования:",
+ "RVC Model Path": "Путь к модели RVC:",
+ "Read from model": "Read from model",
+ "Refresh voice list and index path": "Обновить список голосов и индексов",
+ "Reload device list": "Обновить список устройств",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Изменить частоту дискретизации в выходном файле на финальную. Поставьте 0, чтобы ничего не изменялось:",
+ "Response threshold": "Порог ответа",
+ "Sample length": "Длина сэмпла",
+ "Sampling rate": "Sampling rate",
+ "Save a small final model to the 'weights' folder at each save point": "Сохранять маленькую финальную модель в папку 'weights' на каждой точке сохранения:",
+ "Save file name (default: same as the source file)": "Название сохранённого файла (по умолчанию: такое же, как и у входного):",
+ "Save frequency (save_every_epoch)": "Частота сохранения (save_every_epoch):",
+ "Save name": "Имя файла для сохранения:",
+ "Save only the latest '.ckpt' file to save disk space": "Сохранять только последний файл '.ckpt', чтобы сохранить место на диске:",
+ "Saved model name (without extension)": "Имя файла модели для сохранения (без расширения):",
+ "Sealing date": "Sealing date",
+ "Select Speaker/Singer ID": "Номер говорящего:",
+ "Select the .index file": "Выбрать файл .index",
+ "Select the .pth file": "Выбрать файл .pth",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Выберите алгоритм оценки высоты голоса ('pm': работает быстро, но даёт низкое качество речи; 'harvest': басы лучше, но работает очень медленно; 'crepe': лучшее качество, но сильно нагружает GPU; 'rmvpe': лучшее качество и минимальная нагрузка на GPU):",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "Single inference",
+ "Specify output folder": "Папка для результатов:",
+ "Specify the output folder for accompaniment": "Путь к папке для сохранения аккомпанемента:",
+ "Specify the output folder for vocals": "Путь к папке для сохранения вокала:",
+ "Start audio conversion": "Начать конвертацию аудио",
+ "Step 1: Processing data": "Шаг 1. Переработка данных",
+ "Step 3a: Model training started": "Шаг 3. Запуск обучения модели",
+ "Stop audio conversion": "Закончить конвертацию аудио",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "Частота дискретизации аудио:",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "Это программное обеспечение с открытым исходным кодом распространяется по лицензии MIT. Автор никак не контролирует это программное обеспечение. Пользователи, которые используют эту программу и распространяют аудиозаписи, полученные с помощью этой программы, несут полную ответственность за это. Если вы не согласны с этим, вы не можете использовать какие-либо коды и файлы в рамках этой программы или ссылаться на них. Подробнее в файле Agreement-LICENSE.txt в корневом каталоге программы.",
+ "Total training epochs (total_epoch)": "Полное количество эпох (total_epoch):",
+ "Train": "Обучение модели",
+ "Train feature index": "Обучить индекс черт",
+ "Train model": "Обучить модель",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Обучение модели завершено. Журнал обучения можно просмотреть в консоли или в файле 'train.log' в папке с моделью.",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Изменить высоту голоса (укажите количество полутонов; чтобы поднять голос на октаву, выберите 12, понизить на октаву — -12):",
+ "Unfortunately, there is no compatible GPU available to support your training.": "К сожалению, у вас нету графического процессора, который поддерживает обучение моделей.",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "Выгрузить модель из памяти GPU для освобождения ресурсов",
+ "Version": "Версия архитектуры модели:",
+ "View": "Просмотреть информацию",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Разделение вокала/аккомпанемента и удаление эхо",
+ "Weight (w) for Model A": "Весы (w) модели А:",
+ "Whether the model has pitch guidance": "Поддерживает ли модель изменение высоты голоса (1: да, 0: нет):",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Поддерживает ли модель изменение высоты голоса (1: да, 0: нет):",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Поддержка изменения высоты звука (обязательно для пения, необязательно для речи):",
+ "Yes": "Да",
+ "ckpt Processing": "Обработка ckpt",
+ "index path cannot contain unicode characters": "Путь к файлу индекса",
+ "pth path cannot contain unicode characters": "Путь к файлу pth",
+ "step2:Pitch extraction & feature extraction": "step2:Pitch extraction & feature extraction"
+}
diff --git a/i18n/locale/tr_TR.json b/i18n/locale/tr_TR.json
new file mode 100644
index 0000000000000000000000000000000000000000..961041014878aedc217eba2ec13cfb142f5d52cf
--- /dev/null
+++ b/i18n/locale/tr_TR.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### Model çıkartma\n> Büyük dosya modeli yolunu 'logs' klasöründe girin.\n\nBu, eğitimi yarıda bırakmak istediğinizde ve manuel olarak küçük bir model dosyası çıkartmak ve kaydetmek istediğinizde veya bir ara modeli test etmek istediğinizde kullanışlıdır.",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Model bilgilerini düzenle\n> Sadece 'weights' klasöründen çıkarılan küçük model dosyaları desteklenir.",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Adım 1. Deneysel yapılandırmayı doldurun.\nDeneysel veriler 'logs' klasöründe saklanır ve her bir deney için ayrı bir klasör vardır. Deneysel adı yolu manuel olarak girin; bu yol, deneysel yapılandırmayı, günlükleri ve eğitilmiş model dosyalarını içerir.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### Step 3. Start training.\nFill in the training settings and start training the model and index.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### Model bilgilerini görüntüle\n> Sadece 'weights' klasöründen çıkarılan küçük model dosyaları desteklenir.",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).",
+ "Actually calculated": "Actually calculated",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "Sesin hacim zarfını ayarlayın. 0'a yakın değerler, sesin orijinal vokallerin hacmine benzer olmasını sağlar. Düşük bir değerle ses gürültüsünü maskeleyebilir ve hacmi daha doğal bir şekilde duyulabilir hale getirebilirsiniz. 1'e yaklaştıkça sürekli bir yüksek ses seviyesi elde edilir:",
+ "Algorithmic delays (ms)": "算法延迟(ms)",
+ "All processes have been completed!": "Tüm işlemler tamamlandı!",
+ "Audio device": "Ses cihazı",
+ "Auto-detect index path and select from the dropdown": "İndeks yolunu otomatik olarak tespit et ve açılır menüden seçim yap.",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "Toplu dönüştür. Dönüştürülecek ses dosyalarının bulunduğu klasörü girin veya birden çok ses dosyasını yükleyin. Dönüştürülen ses dosyaları belirtilen klasöre ('opt' varsayılan olarak) dönüştürülecektir",
+ "Batch inference": "Batch inference",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "Batch işleme kullanarak vokal eşlik ayrımı için UVR5 modeli kullanılır.
Geçerli bir klasör yol formatı örneği: D:\\path\\to\\input\\folder (dosya yöneticisi adres çubuğundan kopyalanır).
Model üç kategoriye ayrılır:
1. Vokalleri koru: Bu seçeneği, harmoni içermeyen sesler için kullanın. HP5'ten daha iyi bir şekilde vokalleri korur. İki dahili model içerir: HP2 ve HP3. HP3, eşlik sesini hafifçe sızdırabilir, ancak vokalleri HP2'den biraz daha iyi korur.
2. Sadece ana vokalleri koru: Bu seçeneği, harmoni içeren sesler için kullanın. Ana vokalleri zayıflatabilir. Bir dahili model içerir: HP5.
3. Reverb ve gecikme modelleri (FoxJoy tarafından):
(1) MDX-Net: Stereo reverb'i kaldırmak için en iyi seçenek, ancak mono reverb'i kaldıramaz;
(234) DeEcho: Gecikme efektlerini kaldırır. Agresif mod, Normal moda göre daha kapsamlı bir şekilde kaldırma yapar. DeReverb ayrıca reverb'i kaldırır ve mono reverb'i kaldırabilir, ancak yoğun yankılı yüksek frekanslı içerikler için çok etkili değildir.
Reverb/gecikme notları:
1. DeEcho-DeReverb modelinin işleme süresi diğer iki DeEcho modeline göre yaklaşık olarak iki kat daha uzundur.
2. MDX-Net-Dereverb modeli oldukça yavaştır.
3. Tavsiye edilen en temiz yapılandırma önce MDX-Net'i uygulamak ve ardından DeEcho-Aggressive uygulamaktır.",
+ "Batch size per GPU": "Her GPU için yığın boyutu (batch_size):",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "Tüm eğitim verilerini GPU belleğine önbelleğe alıp almayacağınızı belirtin. Küçük veri setlerini (10 dakikadan az) önbelleğe almak eğitimi hızlandırabilir, ancak büyük veri setlerini önbelleğe almak çok fazla GPU belleği tüketir ve çok fazla hız artışı sağlamaz:",
+ "Calculate": "Calculate",
+ "Choose sample rate of the device": "Choose sample rate of the device",
+ "Choose sample rate of the model": "Choose sample rate of the model",
+ "Convert": "Dönüştür",
+ "Device type": "Device type",
+ "Enable phase vocoder": "Enable phase vocoder",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "GPU indekslerini '-' ile ayırarak girin, örneğin 0-1-2, GPU 0, 1 ve 2'yi kullanmak için:",
+ "Enter the experiment name": "Deneysel adı girin:",
+ "Enter the path of the audio folder to be processed": "İşlenecek ses klasörünün yolunu girin:",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "İşlenecek ses klasörünün yolunu girin (dosya yöneticisinin adres çubuğundan kopyalayın):",
+ "Enter the path of the training folder": "Eğitim klasörünün yolunu girin:",
+ "Exist": "Exist",
+ "Export Onnx": "Onnx Dışa Aktar",
+ "Export Onnx Model": "Onnx Modeli Dışa Aktar",
+ "Export audio (click on the three dots in the lower right corner to download)": "Ses dosyasını dışa aktar (indirmek için sağ alt köşedeki üç noktaya tıklayın)",
+ "Export file format": "Dışa aktarma dosya formatı",
+ "Extra inference time": "Ekstra çıkartma süresi",
+ "Extract": "Çıkart",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0 eğrisi dosyası (isteğe bağlı). Her satırda bir pitch değeri bulunur. Varsayılan F0 ve pitch modülasyonunu değiştirir:",
+ "FAQ (Frequently Asked Questions)": "Sıkça Sorulan Sorular (SSS)",
+ "Fade length": "Geçiş (Fade) uzunluğu",
+ "Fail": "Fail",
+ "Feature extraction": "Özellik çıkartma",
+ "Feature searching ratio": "Arama özelliği oranı (vurgu gücünü kontrol eder, çok yüksek olması sanal etkilere neden olur)",
+ "Formant offset": "Formant offset",
+ "Fusion": "Birleştir",
+ "GPU Information": "GPU Bilgisi",
+ "General settings": "Genel ayarlar",
+ "Hidden": "Hidden",
+ "ID of model A (long)": "ID of model A (long)",
+ "ID of model B (long)": "ID of model B (long)",
+ "ID(long)": "ID (long)",
+ "ID(short)": "ID (short)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": "Eğer >=3 ise, elde edilen pitch sonuçlarına median filtreleme uygula. Bu değer, filtre yarıçapını temsil eder ve nefesliliği azaltabilir.",
+ "Inference time (ms)": "Çıkarsama süresi (ms)",
+ "Inferencing voice": "Ses çıkartma (Inference):",
+ "Information": "Information",
+ "Input device": "Giriş cihazı",
+ "Input noise reduction": "Giriş gürültü azaltma",
+ "Input voice monitor": "Input voice monitor",
+ "Link index to outside folder": "Link index to outside folder",
+ "Load model": "Model yükle",
+ "Load pre-trained base model D path": "Önceden eğitilmiş temel D modelini yükleme yolu:",
+ "Load pre-trained base model G path": "Önceden eğitilmiş temel G modelini yükleme yolu:",
+ "Loudness factor": "ses yüksekliği faktörü",
+ "Model": "Model",
+ "Model Author": "Model Author",
+ "Model Author (Nullable)": "Model Author (Nullable)",
+ "Model Inference": "Model çıkartma (Inference)",
+ "Model architecture version": "Model mimari versiyonu:",
+ "Model info": "Model info",
+ "Model information to be modified": "Düzenlenecek model bilgileri:",
+ "Model information to be placed": "Eklemek için model bilgileri:",
+ "Model name": "Model name",
+ "Modify": "Düzenle",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "Ses dosyaları ayrıca toplu olarak, iki seçimle, öncelikli okuma klasörüyle içe aktarılabilir",
+ "No": "Hayır",
+ "None": "None",
+ "Not exist": "Not exist",
+ "Number of CPU processes used for harvest pitch algorithm": "Number of CPU processes used for harvest pitch algorithm",
+ "Number of CPU processes used for pitch extraction and data processing": "Ses yüksekliği çıkartmak (Pitch) ve verileri işlemek için kullanılacak CPU işlemci sayısı:",
+ "One-click training": "Tek Tuşla Eğit",
+ "Onnx Export Path": "Onnx Dışa Aktarım Yolu:",
+ "Output converted voice": "输出变声",
+ "Output device": "Çıkış cihazı",
+ "Output information": "Çıkış bilgisi",
+ "Output noise reduction": "Çıkış gürültü azaltma",
+ "Path to Model": "Model Yolu:",
+ "Path to Model A": "A Modeli Yolu:",
+ "Path to Model B": "B Modeli Yolu:",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "Özellik indeksi dosyasının yolunu belirtin. Seçilen sonucu kullanmak için boş bırakın veya açılır menüden seçim yapın.",
+ "Performance settings": "Performans ayarları",
+ "Pitch detection algorithm": "Pitch detection algorithm",
+ "Pitch guidance (f0)": "Pitch guidance (f0)",
+ "Pitch settings": "Pitch ayarları",
+ "Please choose the .index file": "Lütfen .index dosyası seçin",
+ "Please choose the .pth file": "Lütfen .pth dosyası seçin",
+ "Please specify the speaker/singer ID": "Lütfen konuşmacı/sanatçı no belirtin:",
+ "Process data": "Verileri işle",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "Sessiz ünsüzleri ve nefes seslerini koruyarak elektronik müzikte yırtılma gibi sanal hataların oluşmasını engeller. 0.5 olarak ayarlandığında devre dışı kalır. Değerin azaltılması korumayı artırabilir, ancak indeksleme doğruluğunu azaltabilir:",
+ "RVC Model Path": "RVC Model Yolu:",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "Ses listesini ve indeks yolunu yenile",
+ "Reload device list": "Cihaz listesini yeniden yükle",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "Son işleme aşamasında çıktı sesini son örnekleme hızına yeniden örnekle. 0 değeri için yeniden örnekleme yapılmaz:",
+ "Response threshold": "Tepki eşiği",
+ "Sample length": "Örnekleme uzunluğu",
+ "Sampling rate": "Sampling rate",
+ "Save a small final model to the 'weights' folder at each save point": "Her kaydetme noktasında son küçük bir modeli 'weights' klasörüne kaydetmek için:",
+ "Save file name (default: same as the source file)": "Kaydedilecek dosya adı (varsayılan: kaynak dosya ile aynı):",
+ "Save frequency (save_every_epoch)": "Kaydetme sıklığı (save_every_epoch):",
+ "Save name": "Kaydetme Adı:",
+ "Save only the latest '.ckpt' file to save disk space": "Sadece en son '.ckpt' dosyasını kaydet:",
+ "Saved model name (without extension)": "Kaydedilecek model adı (uzantı olmadan):",
+ "Sealing date": "Sealing date",
+ "Select Speaker/Singer ID": "Konuşmacı/Şarkıcı No seçin:",
+ "Select the .index file": ".index dosyası seç",
+ "Select the .pth file": ".pth dosyası seç",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "Pitch algoritmasını seçin ('pm': daha hızlı çıkarır ancak daha düşük kaliteli konuşma; 'harvest': daha iyi konuşma sesi ancak son derece yavaş; 'crepe': daha da iyi kalite ancak GPU yoğunluğu gerektirir):",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU",
+ "Similarity": "Similarity",
+ "Similarity (from 0 to 1)": "Similarity (from 0 to 1)",
+ "Single inference": "Single inference",
+ "Specify output folder": "Çıkış klasörünü belirt:",
+ "Specify the output folder for accompaniment": "Müzik ve diğer sesler için çıkış klasörünü belirtin:",
+ "Specify the output folder for vocals": "Vokal için çıkış klasörünü belirtin:",
+ "Start audio conversion": "Ses dönüştürmeyi başlat",
+ "Step 1: Processing data": "Adım 1: Veri işleme",
+ "Step 3a: Model training started": "Adım 3a: Model eğitimi başladı",
+ "Stop audio conversion": "Ses dönüştürmeyi durdur",
+ "Successfully built index into": "Successfully built index into",
+ "Takeover WASAPI device": "Takeover WASAPI device",
+ "Target sample rate": "Hedef örnekleme oranı:",
+ "The audio file to be processed": "The audio file to be processed",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "Bu yazılım, MIT lisansı altında açık kaynaklıdır. Yazarın yazılım üzerinde herhangi bir kontrolü yoktur. Yazılımı kullanan ve yazılım tarafından dışa aktarılan sesleri dağıtan kullanıcılar sorumludur.
Eğer bu maddeyle aynı fikirde değilseniz, yazılım paketi içindeki herhangi bir kod veya dosyayı kullanamaz veya referans göremezsiniz. Detaylar için kök dizindeki Agreement-LICENSE.txt dosyasına bakınız.",
+ "Total training epochs (total_epoch)": "Toplam eğitim turu (total_epoch):",
+ "Train": "Eğitim",
+ "Train feature index": "Özellik Dizinini Eğit",
+ "Train model": "Modeli Eğit",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "Eğitim tamamlandı. Eğitim günlüklerini konsolda veya deney klasörü altındaki train.log dosyasında kontrol edebilirsiniz.",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "Transpoze et (tamsayı, yarıton sayısıyla; bir oktav yükseltmek için: 12, bir oktav düşürmek için: -12):",
+ "Unfortunately, there is no compatible GPU available to support your training.": "Maalesef, eğitiminizi desteklemek için uyumlu bir GPU bulunmamaktadır.",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "GPU bellek kullanımını azaltmak için sesi kaldır",
+ "Version": "Sürüm",
+ "View": "Görüntüle",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "Vokal/Müzik Ayrıştırma ve Yankı Giderme",
+ "Weight (w) for Model A": "A Modeli Ağırlığı:",
+ "Whether the model has pitch guidance": "Modelin ses yüksekliği rehberi içerip içermediği:",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "Modelin ses yüksekliği rehberi içerip içermediği (1: evet, 0: hayır):",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "Modelin ses yüksekliği (Pitch) rehberliği içerip içermediği (şarkı söyleme için şarttır, konuşma için isteğe bağlıdır):",
+ "Yes": "Evet",
+ "ckpt Processing": "ckpt İşleme",
+ "index path cannot contain unicode characters": ".index dosya yolu Çince karakter içeremez",
+ "pth path cannot contain unicode characters": ".pth dosya yolu Çince karakter içeremez",
+ "step2:Pitch extraction & feature extraction": "step2:Pitch extraction & feature extraction"
+}
diff --git a/i18n/locale/zh_CN.json b/i18n/locale/zh_CN.json
new file mode 100644
index 0000000000000000000000000000000000000000..07cfb983f26d7bd7d34e0d61891f681c1c96e427
--- /dev/null
+++ b/i18n/locale/zh_CN.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### 模型比较\n> 模型ID(长)请于下方`查看模型信息`中获得\n\n可用于比较两模型推理相似度",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### 模型提取\n> 输入logs文件夹下大文件模型路径\n\n适用于训一半不想训了模型没有自动提取保存小文件模型, 或者想测试中间模型的情况",
+ "### Model fusion\nCan be used to test timbre fusion.": "### 模型融合\n可用于测试音色融合",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 修改模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### 第一步 填写实验配置\n实验数据放在logs下, 每个实验一个文件夹, 需手工输入实验名路径, 内含实验配置, 日志, 训练得到的模型文件.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### 第二步 音频处理\n#### 1. 音频切片\n自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### 第三步 开始训练\n填写训练设置, 开始训练模型和索引.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 查看模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. 特征提取\n使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号).",
+ "Actually calculated": "实际计算",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "输入源音量包络替换输出音量包络融合比例，越靠近1越使用输出包络",
+ "Algorithmic delays (ms)": "算法延迟(ms)",
+ "All processes have been completed!": "全流程结束！",
+ "Audio device": "音频设备",
+ "Auto-detect index path and select from the dropdown": "自动检测index路径,下拉式选择(dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "批量转换, 输入待转换音频文件夹, 或上传多个音频文件, 在指定文件夹(默认opt)下输出转换的音频. ",
+ "Batch inference": "批量推理",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "人声伴奏分离批量处理， 使用UVR5模型。
合格的文件夹路径格式举例： E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例(去文件管理器地址栏拷就行了)。
模型分为三类：
1、保留人声：不带和声的音频选这个，对主人声保留比HP5更好。内置HP2和HP3两个模型，HP3可能轻微漏伴奏但对主人声保留比HP2稍微好一丁点；
2、仅保留主人声：带和声的音频选这个，对主人声可能有削弱。内置HP5一个模型；
3、去混响、去延迟模型（by FoxJoy）：
(1)MDX-Net(onnx_dereverb):对于双通道混响是最好的选择，不能去除单通道混响；
(234)DeEcho:去除延迟效果。Aggressive比Normal去除得更彻底，DeReverb额外去除混响，可去除单声道混响，但是对高频重的板式混响去不干净。
去混响/去延迟，附：
1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；
2、MDX-Net-Dereverb模型挺慢的；
3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。",
+ "Batch size per GPU": "每张显卡的batch_size",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "是否缓存所有训练集至显存. 10min以下小数据可缓存以加速训练, 大数据缓存会炸显存也加不了多少速",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "使用设备采样率",
+ "Choose sample rate of the model": "使用模型采样率",
+ "Convert": "转换",
+ "Device type": "设备类型",
+ "Enable phase vocoder": "启用相位声码器",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpe卡号配置：以-分隔输入使用的不同进程卡号,例如0-0-1使用在卡0上跑2个进程并在卡1上跑1个进程",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "以-分隔输入使用的卡号, 例如 0-1-2 使用卡0和卡1和卡2",
+ "Enter the experiment name": "输入实验名",
+ "Enter the path of the audio folder to be processed": "输入待处理音频文件夹路径",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "输入待处理音频文件夹路径(去文件管理器地址栏拷就行了)",
+ "Enter the path of the training folder": "输入训练文件夹路径",
+ "Exist": "有",
+ "Export Onnx": "Onnx导出",
+ "Export Onnx Model": "导出Onnx模型",
+ "Export audio (click on the three dots in the lower right corner to download)": "输出音频(右下角三个点,点了可以下载)",
+ "Export file format": "导出文件格式",
+ "Extra inference time": "额外推理时长",
+ "Extract": "提取",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0曲线文件, 可选, 一行一个音高, 代替默认F0及升降调",
+ "FAQ (Frequently Asked Questions)": "常见问题解答",
+ "Fade length": "淡入淡出长度",
+ "Fail": "失败",
+ "Feature extraction": "特征提取",
+ "Feature searching ratio": "检索特征占比",
+ "Formant offset": "共振偏移",
+ "Fusion": "融合",
+ "GPU Information": "显卡信息",
+ "General settings": "常规设置",
+ "Hidden": "不显示",
+ "ID of model A (long)": "A模型ID(长)",
+ "ID of model B (long)": "B模型ID(长)",
+ "ID(long)": "ID(long)",
+ "ID(short)": "ID(短)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3则使用对harvest音高识别的结果使用中值滤波，数值为滤波半径，使用可以削弱哑音",
+ "Inference time (ms)": "推理时间(ms)",
+ "Inferencing voice": "推理音色",
+ "Information": "信息",
+ "Input device": "输入设备",
+ "Input noise reduction": "输入降噪",
+ "Input voice monitor": "输入监听",
+ "Link index to outside folder": "链接索引到外部",
+ "Load model": "加载模型",
+ "Load pre-trained base model D path": "加载预训练底模D路径",
+ "Load pre-trained base model G path": "加载预训练底模G路径",
+ "Loudness factor": "响度因子",
+ "Model": "模型",
+ "Model Author": "模型作者",
+ "Model Author (Nullable)": "模型作者(可空)",
+ "Model Inference": "模型推理",
+ "Model architecture version": "模型版本型号",
+ "Model info": "模型信息",
+ "Model information to be modified": "要改的模型信息",
+ "Model information to be placed": "要置入的模型信息",
+ "Model name": "模型名",
+ "Modify": "修改",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "也可批量输入音频文件, 二选一, 优先读文件夹",
+ "No": "否",
+ "None": "空",
+ "Not exist": "无",
+ "Number of CPU processes used for harvest pitch algorithm": "harvest进程数",
+ "Number of CPU processes used for pitch extraction and data processing": "提取音高和处理数据使用的CPU进程数",
+ "One-click training": "一键训练",
+ "Onnx Export Path": "Onnx输出路径",
+ "Output converted voice": "输出变声",
+ "Output device": "输出设备",
+ "Output information": "输出信息",
+ "Output noise reduction": "输出降噪",
+ "Path to Model": "模型路径",
+ "Path to Model A": "A模型路径",
+ "Path to Model B": "B模型路径",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "特征检索库文件路径,为空则使用下拉的选择结果",
+ "Performance settings": "性能设置",
+ "Pitch detection algorithm": "音高算法",
+ "Pitch guidance (f0)": "音高引导(f0)",
+ "Pitch settings": "音调设置",
+ "Please choose the .index file": "请选择index文件",
+ "Please choose the .pth file": "请选择pth文件",
+ "Please specify the speaker/singer ID": "请指定说话人id",
+ "Process data": "处理数据",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "保护清辅音和呼吸声，防止电音撕裂等artifact，拉满0.5不开启，调低加大保护力度但可能降低索引效果",
+ "RVC Model Path": "RVC模型路径",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "刷新音色列表和索引路径",
+ "Reload device list": "重载设备列表",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "后处理重采样至最终采样率，0为不进行重采样",
+ "Response threshold": "响应阈值",
+ "Sample length": "采样长度",
+ "Sampling rate": "采样率",
+ "Save a small final model to the 'weights' folder at each save point": "是否在每次保存时间点将最终小模型保存至weights文件夹",
+ "Save file name (default: same as the source file)": "保存的文件名, 默认空为和源文件同名",
+ "Save frequency (save_every_epoch)": "保存频率save_every_epoch",
+ "Save name": "保存名",
+ "Save only the latest '.ckpt' file to save disk space": "是否仅保存最新的ckpt文件以节省硬盘空间",
+ "Saved model name (without extension)": "保存的模型名不带后缀",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "请选择说话人id",
+ "Select the .index file": "选择.index文件",
+ "Select the .pth file": "选择.pth文件",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "选择音高提取算法,输入歌声可用pm提速,harvest低音好但巨慢无比,crepe效果好但吃GPU,rmvpe效果最好且微吃GPU",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "选择音高提取算法:输入歌声可用pm提速,高质量语音但CPU差可用dio提速,harvest质量更好但慢,rmvpe效果最好且微吃CPU/GPU",
+ "Similarity": "相似度",
+ "Similarity (from 0 to 1)": "相似度(0到1)",
+ "Single inference": "单次推理",
+ "Specify output folder": "指定输出文件夹",
+ "Specify the output folder for accompaniment": "指定输出非主人声文件夹",
+ "Specify the output folder for vocals": "指定输出主人声文件夹",
+ "Start audio conversion": "开始音频转换",
+ "Step 1: Processing data": "step1:正在处理数据",
+ "Step 3a: Model training started": "step3a:正在训练模型",
+ "Stop audio conversion": "停止音频转换",
+ "Successfully built index into": "成功构建索引到",
+ "Takeover WASAPI device": "独占 WASAPI 设备",
+ "Target sample rate": "目标采样率",
+ "The audio file to be processed": "待处理音频文件",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "本软件以MIT协议开源, 作者不对软件具备任何控制力, 使用软件者、传播软件导出的声音者自负全责.
如不认可该条款, 则不能使用或引用软件包内任何代码和文件. 详见根目录LICENSE.",
+ "Total training epochs (total_epoch)": "总训练轮数total_epoch",
+ "Train": "训练",
+ "Train feature index": "训练特征索引",
+ "Train model": "训练模型",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "变调(整数, 半音数量, 升八度12降八度-12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "很遗憾您这没有能用的显卡来支持您训练",
+ "Unknown": "未知",
+ "Unload model to save GPU memory": "卸载音色省显存",
+ "Version": "版本",
+ "View": "查看",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "伴奏人声分离&去混响&去回声",
+ "Weight (w) for Model A": "A模型权重",
+ "Whether the model has pitch guidance": "模型是否带音高指导",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "模型是否带音高指导,1是0否",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "模型是否带音高指导(唱歌一定要, 语音可以不要)",
+ "Yes": "是",
+ "ckpt Processing": "ckpt处理",
+ "index path cannot contain unicode characters": "index文件路径不可包含中文",
+ "pth path cannot contain unicode characters": "pth文件路径不可包含中文",
+ "step2:Pitch extraction & feature extraction": "step2:正在提取音高&正在提取特征"
+}
diff --git a/i18n/locale/zh_HK.json b/i18n/locale/zh_HK.json
new file mode 100644
index 0000000000000000000000000000000000000000..e996ab75e4aa58f799d6ee26c58a5922f7cb34b1
--- /dev/null
+++ b/i18n/locale/zh_HK.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### 模型提取\n> 输入logs文件夹下大文件模型路径\n\n适用于训一半不想训了模型没有自动提取保存小文件模型, 或者想测试中间模型的情况",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 修改模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### 第二步 音频处理\n#### 1. 音频切片\n自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### 第三步 开始训练\n填写训练设置, 开始训练模型和索引.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 查看模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. 特征提取\n使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号).",
+ "Actually calculated": "实际计算",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "輸入源音量包絡替換輸出音量包絡融合比例，越靠近1越使用輸出包絡",
+ "Algorithmic delays (ms)": "算法延遲(ms)",
+ "All processes have been completed!": "全流程结束！",
+ "Audio device": "音訊設備",
+ "Auto-detect index path and select from the dropdown": "自動檢測index路徑,下拉式選擇(dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "批量轉換，輸入待轉換音頻資料夾，或上傳多個音頻檔案，在指定資料夾(默認opt)下輸出轉換的音頻。",
+ "Batch inference": "批量推理",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "使用UVR5模型進行人聲伴奏分離的批次處理。
有效資料夾路徑格式的例子：D:\\path\\to\\input\\folder（從檔案管理員地址欄複製）。
模型分為三類：
1. 保留人聲：選擇這個選項適用於沒有和聲的音訊。它比HP5更好地保留了人聲。它包括兩個內建模型：HP2和HP3。HP3可能輕微漏出伴奏，但比HP2更好地保留了人聲；
2. 僅保留主人聲：選擇這個選項適用於有和聲的音訊。它可能會削弱主人聲。它包括一個內建模型：HP5。
3. 消除混響和延遲模型（由FoxJoy提供）：
(1) MDX-Net：對於立體聲混響的移除是最好的選擇，但不能移除單聲道混響；
(234) DeEcho：移除延遲效果。Aggressive模式比Normal模式移除得更徹底。DeReverb另外移除混響，可以移除單聲道混響，但對於高頻重的板式混響移除不乾淨。
消除混響/延遲注意事項：
1. DeEcho-DeReverb模型的處理時間是其他兩個DeEcho模型的近兩倍；
2. MDX-Net-Dereverb模型相當慢；
3. 個人推薦的最乾淨配置是先使用MDX-Net，然後使用DeEcho-Aggressive。",
+ "Batch size per GPU": "每张显卡的batch_size",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "是否緩存所有訓練集至 VRAM。小於10分鐘的小數據可緩存以加速訓練，大數據緩存會爆 VRAM 也加不了多少速度",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "使用设备采样率",
+ "Choose sample rate of the model": "使用模型采样率",
+ "Convert": "轉換",
+ "Device type": "设备类型",
+ "Enable phase vocoder": "启用相位声码器",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpe卡號配置：以-分隔輸入使用的不同進程卡號,例如0-0-1使用在卡0上跑2個進程並在卡1上跑1個進程",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "以-分隔輸入使用的卡號, 例如 0-1-2 使用卡0和卡1和卡2",
+ "Enter the experiment name": "輸入實驗名稱",
+ "Enter the path of the audio folder to be processed": "輸入待處理音頻資料夾路徑",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "輸入待處理音頻資料夾路徑(去檔案管理器地址欄拷貝即可)",
+ "Enter the path of the training folder": "輸入訓練檔案夾路徑",
+ "Exist": "有",
+ "Export Onnx": "Onnx导出",
+ "Export Onnx Model": "导出Onnx模型",
+ "Export audio (click on the three dots in the lower right corner to download)": "輸出音頻(右下角三個點，點了可以下載)",
+ "Export file format": "導出檔格式",
+ "Extra inference time": "額外推理時長",
+ "Extract": "提取",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0曲線檔案，可選，一行一個音高，代替預設的F0及升降調",
+ "FAQ (Frequently Asked Questions)": "常見問題解答",
+ "Fade length": "淡入淡出長度",
+ "Fail": "失败",
+ "Feature extraction": "特徵提取",
+ "Feature searching ratio": "檢索特徵佔比",
+ "Formant offset": "共振偏移",
+ "Fusion": "融合",
+ "GPU Information": "顯示卡資訊",
+ "General settings": "一般設定",
+ "Hidden": "不显示",
+ "ID of model A (long)": "A模型ID(长)",
+ "ID of model B (long)": "B模型ID(长)",
+ "ID(long)": "ID(long)",
+ "ID(short)": "ID(短)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3則使用對harvest音高識別的結果使用中值濾波，數值為濾波半徑，使用可以削弱啞音",
+ "Inference time (ms)": "推理時間(ms)",
+ "Inferencing voice": "推理音色",
+ "Information": "信息",
+ "Input device": "輸入設備",
+ "Input noise reduction": "輸入降噪",
+ "Input voice monitor": "输入监听",
+ "Link index to outside folder": "链接索引到外部",
+ "Load model": "載入模型",
+ "Load pre-trained base model D path": "加載預訓練底模D路徑",
+ "Load pre-trained base model G path": "加載預訓練底模G路徑",
+ "Loudness factor": "響度因子",
+ "Model": "模型",
+ "Model Author": "模型作者",
+ "Model Author (Nullable)": "模型作者(可空)",
+ "Model Inference": "模型推理",
+ "Model architecture version": "模型版本型號",
+ "Model info": "模型信息",
+ "Model information to be modified": "要改的模型資訊",
+ "Model information to be placed": "要置入的模型資訊",
+ "Model name": "模型名",
+ "Modify": "修改",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "也可批量输入音频文件, 二选一, 优先读文件夹",
+ "No": "否",
+ "None": "None",
+ "Not exist": "无",
+ "Number of CPU processes used for harvest pitch algorithm": "harvest進程數",
+ "Number of CPU processes used for pitch extraction and data processing": "提取音高和處理數據使用的CPU進程數",
+ "One-click training": "一鍵訓練",
+ "Onnx Export Path": "Onnx输出路径",
+ "Output converted voice": "输出变声",
+ "Output device": "輸出設備",
+ "Output information": "輸出訊息",
+ "Output noise reduction": "輸出降噪",
+ "Path to Model": "模型路徑",
+ "Path to Model A": "A模型路徑",
+ "Path to Model B": "B模型路徑",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "特徵檢索庫檔路徑,為空則使用下拉的選擇結果",
+ "Performance settings": "效能設定",
+ "Pitch detection algorithm": "音高演算法",
+ "Pitch guidance (f0)": "音高引导(f0)",
+ "Pitch settings": "音調設定",
+ "Please choose the .index file": "请选择index文件",
+ "Please choose the .pth file": "请选择pth文件",
+ "Please specify the speaker/singer ID": "請指定說話人id",
+ "Process data": "處理資料",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "保護清輔音和呼吸聲，防止電音撕裂等artifact，拉滿0.5不開啟，調低加大保護力度但可能降低索引效果",
+ "RVC Model Path": "RVC模型路径",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "刷新音色列表和索引路徑",
+ "Reload device list": "重載設備列表",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "後處理重採樣至最終採樣率，0為不進行重採樣",
+ "Response threshold": "響應閾值",
+ "Sample length": "取樣長度",
+ "Sampling rate": "采样率",
+ "Save a small final model to the 'weights' folder at each save point": "是否在每次保存時間點將最終小模型保存至weights檔夾",
+ "Save file name (default: same as the source file)": "儲存的檔案名，預設空為與來源檔案同名",
+ "Save frequency (save_every_epoch)": "保存頻率save_every_epoch",
+ "Save name": "儲存名",
+ "Save only the latest '.ckpt' file to save disk space": "是否僅保存最新的ckpt檔案以節省硬碟空間",
+ "Saved model name (without extension)": "儲存的模型名不帶副檔名",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "請選擇說話人ID",
+ "Select the .index file": "選擇 .index 檔案",
+ "Select the .pth file": "選擇 .pth 檔案",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "選擇音高提取演算法,輸入歌聲可用pm提速,harvest低音好但巨慢無比,crepe效果好但吃GPU,rmvpe效果最好且微吃GPU",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "选择音高提取算法:输入歌声可用pm提速,高质量语音但CPU差可用dio提速,harvest质量更好但慢,rmvpe效果最好且微吃CPU/GPU",
+ "Similarity": "相似度",
+ "Similarity (from 0 to 1)": "相似度(0到1)",
+ "Single inference": "单次推理",
+ "Specify output folder": "指定輸出資料夾",
+ "Specify the output folder for accompaniment": "指定输出非主人声文件夹",
+ "Specify the output folder for vocals": "指定输出主人声文件夹",
+ "Start audio conversion": "開始音訊轉換",
+ "Step 1: Processing data": "step1:正在处理数据",
+ "Step 3a: Model training started": "step3a:正在训练模型",
+ "Stop audio conversion": "停止音訊轉換",
+ "Successfully built index into": "成功构建索引到",
+ "Takeover WASAPI device": "独占 WASAPI 设备",
+ "Target sample rate": "目標取樣率",
+ "The audio file to be processed": "待处理音频文件",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "本軟體以MIT協議開源，作者不對軟體具備任何控制力，使用軟體者、傳播軟體導出的聲音者自負全責。
如不認可該條款，則不能使用或引用軟體包內任何程式碼和檔案。詳見根目錄使用需遵守的協議-LICENSE.txt。",
+ "Total training epochs (total_epoch)": "總訓練輪數total_epoch",
+ "Train": "訓練",
+ "Train feature index": "訓練特徵索引",
+ "Train model": "訓練模型",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "變調(整數、半音數量、升八度12降八度-12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "很遗憾您这没有能用的显卡来支持您训练",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "卸載音色節省 VRAM",
+ "Version": "版本",
+ "View": "查看",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "伴奏人聲分離&去混響&去回聲",
+ "Weight (w) for Model A": "A模型權重",
+ "Whether the model has pitch guidance": "模型是否帶音高指導",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "模型是否帶音高指導，1是0否",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "模型是否帶音高指導（唱歌一定要，語音可以不要）",
+ "Yes": "是",
+ "ckpt Processing": "ckpt處理",
+ "index path cannot contain unicode characters": "index文件路径不可包含中文",
+ "pth path cannot contain unicode characters": "pth文件路径不可包含中文",
+ "step2:Pitch extraction & feature extraction": "step2:正在提取音高&正在提取特征"
+}
diff --git a/i18n/locale/zh_SG.json b/i18n/locale/zh_SG.json
new file mode 100644
index 0000000000000000000000000000000000000000..e996ab75e4aa58f799d6ee26c58a5922f7cb34b1
--- /dev/null
+++ b/i18n/locale/zh_SG.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### 模型提取\n> 输入logs文件夹下大文件模型路径\n\n适用于训一半不想训了模型没有自动提取保存小文件模型, 或者想测试中间模型的情况",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 修改模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### 第二步 音频处理\n#### 1. 音频切片\n自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### 第三步 开始训练\n填写训练设置, 开始训练模型和索引.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 查看模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. 特征提取\n使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号).",
+ "Actually calculated": "实际计算",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "輸入源音量包絡替換輸出音量包絡融合比例，越靠近1越使用輸出包絡",
+ "Algorithmic delays (ms)": "算法延遲(ms)",
+ "All processes have been completed!": "全流程结束！",
+ "Audio device": "音訊設備",
+ "Auto-detect index path and select from the dropdown": "自動檢測index路徑,下拉式選擇(dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "批量轉換，輸入待轉換音頻資料夾，或上傳多個音頻檔案，在指定資料夾(默認opt)下輸出轉換的音頻。",
+ "Batch inference": "批量推理",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "使用UVR5模型進行人聲伴奏分離的批次處理。
有效資料夾路徑格式的例子：D:\\path\\to\\input\\folder（從檔案管理員地址欄複製）。
模型分為三類：
1. 保留人聲：選擇這個選項適用於沒有和聲的音訊。它比HP5更好地保留了人聲。它包括兩個內建模型：HP2和HP3。HP3可能輕微漏出伴奏，但比HP2更好地保留了人聲；
2. 僅保留主人聲：選擇這個選項適用於有和聲的音訊。它可能會削弱主人聲。它包括一個內建模型：HP5。
3. 消除混響和延遲模型（由FoxJoy提供）：
(1) MDX-Net：對於立體聲混響的移除是最好的選擇，但不能移除單聲道混響；
(234) DeEcho：移除延遲效果。Aggressive模式比Normal模式移除得更徹底。DeReverb另外移除混響，可以移除單聲道混響，但對於高頻重的板式混響移除不乾淨。
消除混響/延遲注意事項：
1. DeEcho-DeReverb模型的處理時間是其他兩個DeEcho模型的近兩倍；
2. MDX-Net-Dereverb模型相當慢；
3. 個人推薦的最乾淨配置是先使用MDX-Net，然後使用DeEcho-Aggressive。",
+ "Batch size per GPU": "每张显卡的batch_size",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "是否緩存所有訓練集至 VRAM。小於10分鐘的小數據可緩存以加速訓練，大數據緩存會爆 VRAM 也加不了多少速度",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "使用设备采样率",
+ "Choose sample rate of the model": "使用模型采样率",
+ "Convert": "轉換",
+ "Device type": "设备类型",
+ "Enable phase vocoder": "启用相位声码器",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpe卡號配置：以-分隔輸入使用的不同進程卡號,例如0-0-1使用在卡0上跑2個進程並在卡1上跑1個進程",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "以-分隔輸入使用的卡號, 例如 0-1-2 使用卡0和卡1和卡2",
+ "Enter the experiment name": "輸入實驗名稱",
+ "Enter the path of the audio folder to be processed": "輸入待處理音頻資料夾路徑",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "輸入待處理音頻資料夾路徑(去檔案管理器地址欄拷貝即可)",
+ "Enter the path of the training folder": "輸入訓練檔案夾路徑",
+ "Exist": "有",
+ "Export Onnx": "Onnx导出",
+ "Export Onnx Model": "导出Onnx模型",
+ "Export audio (click on the three dots in the lower right corner to download)": "輸出音頻(右下角三個點，點了可以下載)",
+ "Export file format": "導出檔格式",
+ "Extra inference time": "額外推理時長",
+ "Extract": "提取",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0曲線檔案，可選，一行一個音高，代替預設的F0及升降調",
+ "FAQ (Frequently Asked Questions)": "常見問題解答",
+ "Fade length": "淡入淡出長度",
+ "Fail": "失败",
+ "Feature extraction": "特徵提取",
+ "Feature searching ratio": "檢索特徵佔比",
+ "Formant offset": "共振偏移",
+ "Fusion": "融合",
+ "GPU Information": "顯示卡資訊",
+ "General settings": "一般設定",
+ "Hidden": "不显示",
+ "ID of model A (long)": "A模型ID(长)",
+ "ID of model B (long)": "B模型ID(长)",
+ "ID(long)": "ID(long)",
+ "ID(short)": "ID(短)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3則使用對harvest音高識別的結果使用中值濾波，數值為濾波半徑，使用可以削弱啞音",
+ "Inference time (ms)": "推理時間(ms)",
+ "Inferencing voice": "推理音色",
+ "Information": "信息",
+ "Input device": "輸入設備",
+ "Input noise reduction": "輸入降噪",
+ "Input voice monitor": "输入监听",
+ "Link index to outside folder": "链接索引到外部",
+ "Load model": "載入模型",
+ "Load pre-trained base model D path": "加載預訓練底模D路徑",
+ "Load pre-trained base model G path": "加載預訓練底模G路徑",
+ "Loudness factor": "響度因子",
+ "Model": "模型",
+ "Model Author": "模型作者",
+ "Model Author (Nullable)": "模型作者(可空)",
+ "Model Inference": "模型推理",
+ "Model architecture version": "模型版本型號",
+ "Model info": "模型信息",
+ "Model information to be modified": "要改的模型資訊",
+ "Model information to be placed": "要置入的模型資訊",
+ "Model name": "模型名",
+ "Modify": "修改",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "也可批量输入音频文件, 二选一, 优先读文件夹",
+ "No": "否",
+ "None": "None",
+ "Not exist": "无",
+ "Number of CPU processes used for harvest pitch algorithm": "harvest進程數",
+ "Number of CPU processes used for pitch extraction and data processing": "提取音高和處理數據使用的CPU進程數",
+ "One-click training": "一鍵訓練",
+ "Onnx Export Path": "Onnx输出路径",
+ "Output converted voice": "输出变声",
+ "Output device": "輸出設備",
+ "Output information": "輸出訊息",
+ "Output noise reduction": "輸出降噪",
+ "Path to Model": "模型路徑",
+ "Path to Model A": "A模型路徑",
+ "Path to Model B": "B模型路徑",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "特徵檢索庫檔路徑,為空則使用下拉的選擇結果",
+ "Performance settings": "效能設定",
+ "Pitch detection algorithm": "音高演算法",
+ "Pitch guidance (f0)": "音高引导(f0)",
+ "Pitch settings": "音調設定",
+ "Please choose the .index file": "请选择index文件",
+ "Please choose the .pth file": "请选择pth文件",
+ "Please specify the speaker/singer ID": "請指定說話人id",
+ "Process data": "處理資料",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "保護清輔音和呼吸聲，防止電音撕裂等artifact，拉滿0.5不開啟，調低加大保護力度但可能降低索引效果",
+ "RVC Model Path": "RVC模型路径",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "刷新音色列表和索引路徑",
+ "Reload device list": "重載設備列表",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "後處理重採樣至最終採樣率，0為不進行重採樣",
+ "Response threshold": "響應閾值",
+ "Sample length": "取樣長度",
+ "Sampling rate": "采样率",
+ "Save a small final model to the 'weights' folder at each save point": "是否在每次保存時間點將最終小模型保存至weights檔夾",
+ "Save file name (default: same as the source file)": "儲存的檔案名，預設空為與來源檔案同名",
+ "Save frequency (save_every_epoch)": "保存頻率save_every_epoch",
+ "Save name": "儲存名",
+ "Save only the latest '.ckpt' file to save disk space": "是否僅保存最新的ckpt檔案以節省硬碟空間",
+ "Saved model name (without extension)": "儲存的模型名不帶副檔名",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "請選擇說話人ID",
+ "Select the .index file": "選擇 .index 檔案",
+ "Select the .pth file": "選擇 .pth 檔案",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "選擇音高提取演算法,輸入歌聲可用pm提速,harvest低音好但巨慢無比,crepe效果好但吃GPU,rmvpe效果最好且微吃GPU",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "选择音高提取算法:输入歌声可用pm提速,高质量语音但CPU差可用dio提速,harvest质量更好但慢,rmvpe效果最好且微吃CPU/GPU",
+ "Similarity": "相似度",
+ "Similarity (from 0 to 1)": "相似度(0到1)",
+ "Single inference": "单次推理",
+ "Specify output folder": "指定輸出資料夾",
+ "Specify the output folder for accompaniment": "指定输出非主人声文件夹",
+ "Specify the output folder for vocals": "指定输出主人声文件夹",
+ "Start audio conversion": "開始音訊轉換",
+ "Step 1: Processing data": "step1:正在处理数据",
+ "Step 3a: Model training started": "step3a:正在训练模型",
+ "Stop audio conversion": "停止音訊轉換",
+ "Successfully built index into": "成功构建索引到",
+ "Takeover WASAPI device": "独占 WASAPI 设备",
+ "Target sample rate": "目標取樣率",
+ "The audio file to be processed": "待处理音频文件",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "本軟體以MIT協議開源，作者不對軟體具備任何控制力，使用軟體者、傳播軟體導出的聲音者自負全責。
如不認可該條款，則不能使用或引用軟體包內任何程式碼和檔案。詳見根目錄使用需遵守的協議-LICENSE.txt。",
+ "Total training epochs (total_epoch)": "總訓練輪數total_epoch",
+ "Train": "訓練",
+ "Train feature index": "訓練特徵索引",
+ "Train model": "訓練模型",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "變調(整數、半音數量、升八度12降八度-12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "很遗憾您这没有能用的显卡来支持您训练",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "卸載音色節省 VRAM",
+ "Version": "版本",
+ "View": "查看",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "伴奏人聲分離&去混響&去回聲",
+ "Weight (w) for Model A": "A模型權重",
+ "Whether the model has pitch guidance": "模型是否帶音高指導",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "模型是否帶音高指導，1是0否",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "模型是否帶音高指導（唱歌一定要，語音可以不要）",
+ "Yes": "是",
+ "ckpt Processing": "ckpt處理",
+ "index path cannot contain unicode characters": "index文件路径不可包含中文",
+ "pth path cannot contain unicode characters": "pth文件路径不可包含中文",
+ "step2:Pitch extraction & feature extraction": "step2:正在提取音高&正在提取特征"
+}
diff --git a/i18n/locale/zh_TW.json b/i18n/locale/zh_TW.json
new file mode 100644
index 0000000000000000000000000000000000000000..e996ab75e4aa58f799d6ee26c58a5922f7cb34b1
--- /dev/null
+++ b/i18n/locale/zh_TW.json
@@ -0,0 +1,159 @@
+{
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.": "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models.",
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model.": "### 模型提取\n> 输入logs文件夹下大文件模型路径\n\n适用于训一半不想训了模型没有自动提取保存小文件模型, 或者想测试中间模型的情况",
+ "### Model fusion\nCan be used to test timbre fusion.": "### Model fusion\nCan be used to test timbre fusion.",
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 修改模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.": "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files.",
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported.": "### 第二步 音频处理\n#### 1. 音频切片\n自动遍历训练文件夹下所有可解码成音频的文件并进行切片归一化, 在实验目录下生成2个wav文件夹; 暂时只支持单人训练.",
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index.": "### 第三步 开始训练\n填写训练设置, 开始训练模型和索引.",
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder.": "### 查看模型信息\n> 仅支持weights文件夹下提取的小模型文件",
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index).": "#### 2. 特征提取\n使用CPU提取音高(如果模型带音高), 使用GPU提取特征(选择卡号).",
+ "Actually calculated": "实际计算",
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume": "輸入源音量包絡替換輸出音量包絡融合比例，越靠近1越使用輸出包絡",
+ "Algorithmic delays (ms)": "算法延遲(ms)",
+ "All processes have been completed!": "全流程结束！",
+ "Audio device": "音訊設備",
+ "Auto-detect index path and select from the dropdown": "自動檢測index路徑,下拉式選擇(dropdown)",
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt').": "批量轉換，輸入待轉換音頻資料夾，或上傳多個音頻檔案，在指定資料夾(默認opt)下輸出轉換的音頻。",
+ "Batch inference": "批量推理",
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive.": "使用UVR5模型進行人聲伴奏分離的批次處理。
有效資料夾路徑格式的例子：D:\\path\\to\\input\\folder（從檔案管理員地址欄複製）。
模型分為三類：
1. 保留人聲：選擇這個選項適用於沒有和聲的音訊。它比HP5更好地保留了人聲。它包括兩個內建模型：HP2和HP3。HP3可能輕微漏出伴奏，但比HP2更好地保留了人聲；
2. 僅保留主人聲：選擇這個選項適用於有和聲的音訊。它可能會削弱主人聲。它包括一個內建模型：HP5。
3. 消除混響和延遲模型（由FoxJoy提供）：
(1) MDX-Net：對於立體聲混響的移除是最好的選擇，但不能移除單聲道混響；
(234) DeEcho：移除延遲效果。Aggressive模式比Normal模式移除得更徹底。DeReverb另外移除混響，可以移除單聲道混響，但對於高頻重的板式混響移除不乾淨。
消除混響/延遲注意事項：
1. DeEcho-DeReverb模型的處理時間是其他兩個DeEcho模型的近兩倍；
2. MDX-Net-Dereverb模型相當慢；
3. 個人推薦的最乾淨配置是先使用MDX-Net，然後使用DeEcho-Aggressive。",
+ "Batch size per GPU": "每张显卡的batch_size",
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement": "是否緩存所有訓練集至 VRAM。小於10分鐘的小數據可緩存以加速訓練，大數據緩存會爆 VRAM 也加不了多少速度",
+ "Calculate": "计算",
+ "Choose sample rate of the device": "使用设备采样率",
+ "Choose sample rate of the model": "使用模型采样率",
+ "Convert": "轉換",
+ "Device type": "设备类型",
+ "Enable phase vocoder": "启用相位声码器",
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1": "rmvpe卡號配置：以-分隔輸入使用的不同進程卡號,例如0-0-1使用在卡0上跑2個進程並在卡1上跑1個進程",
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2": "以-分隔輸入使用的卡號, 例如 0-1-2 使用卡0和卡1和卡2",
+ "Enter the experiment name": "輸入實驗名稱",
+ "Enter the path of the audio folder to be processed": "輸入待處理音頻資料夾路徑",
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)": "輸入待處理音頻資料夾路徑(去檔案管理器地址欄拷貝即可)",
+ "Enter the path of the training folder": "輸入訓練檔案夾路徑",
+ "Exist": "有",
+ "Export Onnx": "Onnx导出",
+ "Export Onnx Model": "导出Onnx模型",
+ "Export audio (click on the three dots in the lower right corner to download)": "輸出音頻(右下角三個點，點了可以下載)",
+ "Export file format": "導出檔格式",
+ "Extra inference time": "額外推理時長",
+ "Extract": "提取",
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation": "F0曲線檔案，可選，一行一個音高，代替預設的F0及升降調",
+ "FAQ (Frequently Asked Questions)": "常見問題解答",
+ "Fade length": "淡入淡出長度",
+ "Fail": "失败",
+ "Feature extraction": "特徵提取",
+ "Feature searching ratio": "檢索特徵佔比",
+ "Formant offset": "共振偏移",
+ "Fusion": "融合",
+ "GPU Information": "顯示卡資訊",
+ "General settings": "一般設定",
+ "Hidden": "不显示",
+ "ID of model A (long)": "A模型ID(长)",
+ "ID of model B (long)": "B模型ID(长)",
+ "ID(long)": "ID(long)",
+ "ID(short)": "ID(短)",
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness.": ">=3則使用對harvest音高識別的結果使用中值濾波，數值為濾波半徑，使用可以削弱啞音",
+ "Inference time (ms)": "推理時間(ms)",
+ "Inferencing voice": "推理音色",
+ "Information": "信息",
+ "Input device": "輸入設備",
+ "Input noise reduction": "輸入降噪",
+ "Input voice monitor": "输入监听",
+ "Link index to outside folder": "链接索引到外部",
+ "Load model": "載入模型",
+ "Load pre-trained base model D path": "加載預訓練底模D路徑",
+ "Load pre-trained base model G path": "加載預訓練底模G路徑",
+ "Loudness factor": "響度因子",
+ "Model": "模型",
+ "Model Author": "模型作者",
+ "Model Author (Nullable)": "模型作者(可空)",
+ "Model Inference": "模型推理",
+ "Model architecture version": "模型版本型號",
+ "Model info": "模型信息",
+ "Model information to be modified": "要改的模型資訊",
+ "Model information to be placed": "要置入的模型資訊",
+ "Model name": "模型名",
+ "Modify": "修改",
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored.": "也可批量输入音频文件, 二选一, 优先读文件夹",
+ "No": "否",
+ "None": "None",
+ "Not exist": "无",
+ "Number of CPU processes used for harvest pitch algorithm": "harvest進程數",
+ "Number of CPU processes used for pitch extraction and data processing": "提取音高和處理數據使用的CPU進程數",
+ "One-click training": "一鍵訓練",
+ "Onnx Export Path": "Onnx输出路径",
+ "Output converted voice": "输出变声",
+ "Output device": "輸出設備",
+ "Output information": "輸出訊息",
+ "Output noise reduction": "輸出降噪",
+ "Path to Model": "模型路徑",
+ "Path to Model A": "A模型路徑",
+ "Path to Model B": "B模型路徑",
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown": "特徵檢索庫檔路徑,為空則使用下拉的選擇結果",
+ "Performance settings": "效能設定",
+ "Pitch detection algorithm": "音高演算法",
+ "Pitch guidance (f0)": "音高引导(f0)",
+ "Pitch settings": "音調設定",
+ "Please choose the .index file": "请选择index文件",
+ "Please choose the .pth file": "请选择pth文件",
+ "Please specify the speaker/singer ID": "請指定說話人id",
+ "Process data": "處理資料",
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy": "保護清輔音和呼吸聲，防止電音撕裂等artifact，拉滿0.5不開啟，調低加大保護力度但可能降低索引效果",
+ "RVC Model Path": "RVC模型路径",
+ "Read from model": "从模型中读取",
+ "Refresh voice list and index path": "刷新音色列表和索引路徑",
+ "Reload device list": "重載設備列表",
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling": "後處理重採樣至最終採樣率，0為不進行重採樣",
+ "Response threshold": "響應閾值",
+ "Sample length": "取樣長度",
+ "Sampling rate": "采样率",
+ "Save a small final model to the 'weights' folder at each save point": "是否在每次保存時間點將最終小模型保存至weights檔夾",
+ "Save file name (default: same as the source file)": "儲存的檔案名，預設空為與來源檔案同名",
+ "Save frequency (save_every_epoch)": "保存頻率save_every_epoch",
+ "Save name": "儲存名",
+ "Save only the latest '.ckpt' file to save disk space": "是否僅保存最新的ckpt檔案以節省硬碟空間",
+ "Saved model name (without extension)": "儲存的模型名不帶副檔名",
+ "Sealing date": "封装时间",
+ "Select Speaker/Singer ID": "請選擇說話人ID",
+ "Select the .index file": "選擇 .index 檔案",
+ "Select the .pth file": "選擇 .pth 檔案",
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement": "選擇音高提取演算法,輸入歌聲可用pm提速,harvest低音好但巨慢無比,crepe效果好但吃GPU,rmvpe效果最好且微吃GPU",
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU": "选择音高提取算法:输入歌声可用pm提速,高质量语音但CPU差可用dio提速,harvest质量更好但慢,rmvpe效果最好且微吃CPU/GPU",
+ "Similarity": "相似度",
+ "Similarity (from 0 to 1)": "相似度(0到1)",
+ "Single inference": "单次推理",
+ "Specify output folder": "指定輸出資料夾",
+ "Specify the output folder for accompaniment": "指定输出非主人声文件夹",
+ "Specify the output folder for vocals": "指定输出主人声文件夹",
+ "Start audio conversion": "開始音訊轉換",
+ "Step 1: Processing data": "step1:正在处理数据",
+ "Step 3a: Model training started": "step3a:正在训练模型",
+ "Stop audio conversion": "停止音訊轉換",
+ "Successfully built index into": "成功构建索引到",
+ "Takeover WASAPI device": "独占 WASAPI 设备",
+ "Target sample rate": "目標取樣率",
+ "The audio file to be processed": "待处理音频文件",
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details.": "本軟體以MIT協議開源，作者不對軟體具備任何控制力，使用軟體者、傳播軟體導出的聲音者自負全責。
如不認可該條款，則不能使用或引用軟體包內任何程式碼和檔案。詳見根目錄使用需遵守的協議-LICENSE.txt。",
+ "Total training epochs (total_epoch)": "總訓練輪數total_epoch",
+ "Train": "訓練",
+ "Train feature index": "訓練特徵索引",
+ "Train model": "訓練模型",
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder.": "训练结束, 您可查看控制台训练日志或实验文件夹下的train.log",
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)": "變調(整數、半音數量、升八度12降八度-12)",
+ "Unfortunately, there is no compatible GPU available to support your training.": "很遗憾您这没有能用的显卡来支持您训练",
+ "Unknown": "Unknown",
+ "Unload model to save GPU memory": "卸載音色節省 VRAM",
+ "Version": "版本",
+ "View": "查看",
+ "Vocals/Accompaniment Separation & Reverberation Removal": "伴奏人聲分離&去混響&去回聲",
+ "Weight (w) for Model A": "A模型權重",
+ "Whether the model has pitch guidance": "模型是否帶音高指導",
+ "Whether the model has pitch guidance (1: yes, 0: no)": "模型是否帶音高指導，1是0否",
+ "Whether the model has pitch guidance (required for singing, optional for speech)": "模型是否帶音高指導（唱歌一定要，語音可以不要）",
+ "Yes": "是",
+ "ckpt Processing": "ckpt處理",
+ "index path cannot contain unicode characters": "index文件路径不可包含中文",
+ "pth path cannot contain unicode characters": "pth文件路径不可包含中文",
+ "step2:Pitch extraction & feature extraction": "step2:正在提取音高&正在提取特征"
+}
diff --git a/i18n/locale_diff.py b/i18n/locale_diff.py
new file mode 100644
index 0000000000000000000000000000000000000000..640734660b746e36cf1c00ccaf7ab7bc66d7a33b
--- /dev/null
+++ b/i18n/locale_diff.py
@@ -0,0 +1,47 @@
+import json
+import os
+from collections import OrderedDict
+
+# Define the standard file name
+standard_file = "locale/en_US.json"
+
+# Find all JSON files in the directory
+dir_path = "locale/"
+languages = [
+ os.path.join(dir_path, f)
+ for f in os.listdir(dir_path)
+ if f.endswith(".json") and f != standard_file
+]
+
+# Load the standard file
+with open(standard_file, "r", encoding="utf-8") as f:
+ standard_data = json.load(f, object_pairs_hook=OrderedDict)
+
+# Loop through each language file
+for lang_file in languages:
+ # Load the language file
+ with open(lang_file, "r", encoding="utf-8") as f:
+ lang_data = json.load(f, object_pairs_hook=OrderedDict)
+
+ # Find the difference between the language file and the standard file
+ diff = set(standard_data.keys()) - set(lang_data.keys())
+
+ miss = set(lang_data.keys()) - set(standard_data.keys())
+
+ # Add any missing keys to the language file
+ for key in diff:
+ lang_data[key] = key
+
+ # Del any extra keys to the language file
+ for key in miss:
+ del lang_data[key]
+
+ # Sort the keys of the language file to match the order of the standard file
+ lang_data = OrderedDict(
+ sorted(lang_data.items(), key=lambda x: list(standard_data.keys()).index(x[0]))
+ )
+
+ # Save the updated language file
+ with open(lang_file, "w", encoding="utf-8") as f:
+ json.dump(lang_data, f, ensure_ascii=False, indent=4, sort_keys=True)
+ f.write("\n")
diff --git a/i18n/scan_i18n.py b/i18n/scan_i18n.py
new file mode 100644
index 0000000000000000000000000000000000000000..901802f9f963ff1f299e6b51a9c725bd2831f8eb
--- /dev/null
+++ b/i18n/scan_i18n.py
@@ -0,0 +1,66 @@
+import ast
+import glob
+import json
+from collections import OrderedDict
+
+
+def extract_i18n_strings(node):
+ i18n_strings = []
+
+ if (
+ isinstance(node, ast.Call)
+ and isinstance(node.func, ast.Name)
+ and node.func.id == "i18n"
+ ):
+ for arg in node.args:
+ if isinstance(arg, ast.Str):
+ i18n_strings.append(arg.s)
+
+ for child_node in ast.iter_child_nodes(node):
+ i18n_strings.extend(extract_i18n_strings(child_node))
+
+ return i18n_strings
+
+
+# scan the directory for all .py files (recursively)
+# for each file, parse the code into an AST
+# for each AST, extract the i18n strings
+
+strings = []
+for filename in glob.iglob("**/*.py", recursive=True):
+ with open(filename, "r") as f:
+ code = f.read()
+ if "I18nAuto" in code:
+ tree = ast.parse(code)
+ i18n_strings = extract_i18n_strings(tree)
+ print(filename, len(i18n_strings))
+ strings.extend(i18n_strings)
+code_keys = set(strings)
+print()
+print("Total unique:", len(code_keys))
+
+
+standard_file = "i18n/locale/en_US.json"
+with open(standard_file, "r", encoding="utf-8") as f:
+ standard_data = json.load(f, object_pairs_hook=OrderedDict)
+standard_keys = set(standard_data.keys())
+
+# Define the standard file name
+unused_keys = standard_keys - code_keys
+print("Unused keys:", len(unused_keys))
+for unused_key in unused_keys:
+ print("\t", unused_key)
+
+missing_keys = code_keys - standard_keys
+print("Missing keys:", len(missing_keys))
+for missing_key in missing_keys:
+ print("\t", missing_key)
+
+code_keys_dict = OrderedDict()
+for s in strings:
+ code_keys_dict[s] = s
+
+# write back
+with open(standard_file, "w", encoding="utf-8") as f:
+ json.dump(code_keys_dict, f, ensure_ascii=False, indent=4, sort_keys=True)
+ f.write("\n")
diff --git a/infer/lib/audio.py b/infer/lib/audio.py
new file mode 100644
index 0000000000000000000000000000000000000000..3bc0bfb770c9e408c50d564a2a9c7cf036d937a8
--- /dev/null
+++ b/infer/lib/audio.py
@@ -0,0 +1,150 @@
+from io import BufferedWriter, BytesIO
+from pathlib import Path
+from typing import Dict, Tuple
+import os
+
+import numpy as np
+import av
+from av.audio.resampler import AudioResampler
+
+video_format_dict: Dict[str, str] = {
+ "m4a": "mp4",
+}
+
+audio_format_dict: Dict[str, str] = {
+ "ogg": "libvorbis",
+ "mp4": "aac",
+}
+
+
+def wav2(i: BytesIO, o: BufferedWriter, format: str):
+ inp = av.open(i, "r")
+ format = video_format_dict.get(format, format)
+ out = av.open(o, "w", format=format)
+ format = audio_format_dict.get(format, format)
+
+ ostream = out.add_stream(format)
+
+ for frame in inp.decode(audio=0):
+ for p in ostream.encode(frame):
+ out.mux(p)
+
+ for p in ostream.encode(None):
+ out.mux(p)
+
+ out.close()
+ inp.close()
+
+
+def load_audio(file: str, sr: int) -> np.ndarray:
+ if not Path(file).exists():
+ raise FileNotFoundError(f"File not found: {file}")
+
+ try:
+ container = av.open(file)
+ resampler = AudioResampler(format="fltp", layout="mono", rate=sr)
+
+ # Estimated maximum total number of samples to pre-allocate the array
+ # AV stores length in microseconds by default
+ estimated_total_samples = int(container.duration * sr // 1_000_000)
+ decoded_audio = np.zeros(estimated_total_samples + 1, dtype=np.float32)
+
+ offset = 0
+ for frame in container.decode(audio=0):
+ frame.pts = None # Clear presentation timestamp to avoid resampling issues
+ resampled_frames = resampler.resample(frame)
+ for resampled_frame in resampled_frames:
+ frame_data = resampled_frame.to_ndarray()[0]
+ end_index = offset + len(frame_data)
+
+ # Check if decoded_audio has enough space, and resize if necessary
+ if end_index > decoded_audio.shape[0]:
+ decoded_audio = np.resize(decoded_audio, end_index + 1)
+
+ decoded_audio[offset:end_index] = frame_data
+ offset += len(frame_data)
+
+ # Truncate the array to the actual size
+ decoded_audio = decoded_audio[:offset]
+ except Exception as e:
+ raise RuntimeError(f"Failed to load audio: {e}")
+
+ return decoded_audio
+
+
+def downsample_audio(input_path: str, output_path: str, format: str) -> None:
+ if not os.path.exists(input_path):
+ return
+
+ input_container = av.open(input_path)
+ output_container = av.open(output_path, "w")
+
+ # Create a stream in the output container
+ input_stream = input_container.streams.audio[0]
+ output_stream = output_container.add_stream(format)
+
+ output_stream.bit_rate = 128_000 # 128kb/s (equivalent to -q:a 2)
+
+ # Copy packets from the input file to the output file
+ for packet in input_container.demux(input_stream):
+ for frame in packet.decode():
+ for out_packet in output_stream.encode(frame):
+ output_container.mux(out_packet)
+
+ for packet in output_stream.encode():
+ output_container.mux(packet)
+
+ # Close the containers
+ input_container.close()
+ output_container.close()
+
+ try: # Remove the original file
+ os.remove(input_path)
+ except Exception as e:
+ print(f"Failed to remove the original file: {e}")
+
+
+def resample_audio(
+ input_path: str, output_path: str, codec: str, format: str, sr: int, layout: str
+) -> None:
+ if not os.path.exists(input_path):
+ return
+
+ input_container = av.open(input_path)
+ output_container = av.open(output_path, "w")
+
+ # Create a stream in the output container
+ input_stream = input_container.streams.audio[0]
+ output_stream = output_container.add_stream(codec, rate=sr, layout=layout)
+
+ resampler = AudioResampler(format, layout, sr)
+
+ # Copy packets from the input file to the output file
+ for packet in input_container.demux(input_stream):
+ for frame in packet.decode():
+ frame.pts = None # Clear presentation timestamp to avoid resampling issues
+ out_frames = resampler.resample(frame)
+ for out_frame in out_frames:
+ for out_packet in output_stream.encode(out_frame):
+ output_container.mux(out_packet)
+
+ for packet in output_stream.encode():
+ output_container.mux(packet)
+
+ # Close the containers
+ input_container.close()
+ output_container.close()
+
+ try: # Remove the original file
+ os.remove(input_path)
+ except Exception as e:
+ print(f"Failed to remove the original file: {e}")
+
+
+def get_audio_properties(input_path: str) -> Tuple:
+ container = av.open(input_path)
+ audio_stream = next(s for s in container.streams if s.type == "audio")
+ channels = 1 if audio_stream.layout == "mono" else 2
+ rate = audio_stream.base_rate
+ container.close()
+ return channels, rate
diff --git a/infer/lib/rtrvc.py b/infer/lib/rtrvc.py
new file mode 100644
index 0000000000000000000000000000000000000000..b55a69b1bb590ed9e1ad11773e972c9b0d7c825f
--- /dev/null
+++ b/infer/lib/rtrvc.py
@@ -0,0 +1,363 @@
+from io import BytesIO
+import os
+from typing import Union, Literal, Optional
+
+import fairseq
+import faiss
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchaudio.transforms import Resample
+
+from rvc.f0 import PM, Harvest, RMVPE, CRePE, Dio, FCPE
+from rvc.synthesizer import load_synthesizer
+
+
+class RVC:
+ def __init__(
+ self,
+ key: Union[int, float],
+ formant: Union[int, float],
+ pth_path: torch.serialization.FILE_LIKE,
+ index_path: str,
+ index_rate: Union[int, float],
+ n_cpu: int = os.cpu_count(),
+ device: str = "cpu",
+ use_jit: bool = False,
+ is_half: bool = False,
+ is_dml: bool = False,
+ ) -> None:
+ if is_dml:
+
+ def forward_dml(ctx, x, scale):
+ ctx.scale = scale
+ res = x.clone().detach()
+ return res
+
+ fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+
+ self.device = device
+ self.f0_up_key = key
+ self.formant_shift = formant
+ self.sr = 16000 # hubert sampling rate
+ self.window = 160 # hop length
+ self.f0_min = 50
+ self.f0_max = 1100
+ self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+ self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+ self.n_cpu = n_cpu
+ self.use_jit = use_jit
+ self.is_half = is_half
+
+ if index_rate > 0:
+ self.index = faiss.read_index(index_path)
+ self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+
+ self.pth_path = pth_path
+ self.index_path = index_path
+ self.index_rate = index_rate
+
+ self.cache_pitch: torch.Tensor = torch.zeros(
+ 1024, device=self.device, dtype=torch.long
+ )
+ self.cache_pitchf = torch.zeros(1024, device=self.device, dtype=torch.float32)
+
+ self.resample_kernel = {}
+
+ self.f0_methods = {
+ "crepe": self._get_f0_crepe,
+ "rmvpe": self._get_f0_rmvpe,
+ "fcpe": self._get_f0_fcpe,
+ "pm": self._get_f0_pm,
+ "harvest": self._get_f0_harvest,
+ "dio": self._get_f0_dio,
+ }
+
+ models, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+ ["assets/hubert/hubert_base.pt"],
+ suffix="",
+ )
+ hubert_model = models[0]
+ hubert_model = hubert_model.to(self.device)
+ if self.is_half:
+ hubert_model = hubert_model.half()
+ else:
+ hubert_model = hubert_model.float()
+ hubert_model.eval()
+ self.hubert = hubert_model
+
+ self.net_g: Optional[nn.Module] = None
+
+ def set_default_model():
+ self.net_g, cpt = load_synthesizer(self.pth_path, self.device)
+ self.tgt_sr = cpt["config"][-1]
+ cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+ self.if_f0 = cpt.get("f0", 1)
+ self.version = cpt.get("version", "v1")
+ if self.is_half:
+ self.net_g = self.net_g.half()
+ else:
+ self.net_g = self.net_g.float()
+
+ def set_jit_model():
+ from rvc.jit import get_jit_model
+ from rvc.synthesizer import synthesizer_jit_export
+
+ cpt = get_jit_model(self.pth_path, self.is_half, synthesizer_jit_export)
+
+ self.tgt_sr = cpt["config"][-1]
+ self.if_f0 = cpt.get("f0", 1)
+ self.version = cpt.get("version", "v1")
+ self.net_g = torch.jit.load(BytesIO(cpt["model"]), map_location=self.device)
+ self.net_g.infer = self.net_g.forward
+ self.net_g.eval().to(self.device)
+
+ if (
+ self.use_jit
+ and not is_dml
+ and not (self.is_half and "cpu" in str(self.device))
+ ):
+ set_jit_model()
+ else:
+ set_default_model()
+
+ def set_key(self, new_key):
+ self.f0_up_key = new_key
+
+ def set_formant(self, new_formant):
+ self.formant_shift = new_formant
+
+ def set_index_rate(self, new_index_rate):
+ if new_index_rate > 0 and self.index_rate <= 0:
+ self.index = faiss.read_index(self.index_path)
+ self.big_npy = self.index.reconstruct_n(0, self.index.ntotal)
+ self.index_rate = new_index_rate
+
+ def infer(
+ self,
+ input_wav: torch.Tensor,
+ block_frame_16k: int,
+ skip_head: int,
+ return_length: int,
+ f0method: Union[tuple, str],
+ inp_f0: Optional[np.ndarray] = None,
+ protect: float = 1.0,
+ ) -> np.ndarray:
+ with torch.no_grad():
+ if self.is_half:
+ feats = input_wav.half()
+ else:
+ feats = input_wav.float()
+ feats = feats.to(self.device)
+ if feats.dim() == 2: # double channels
+ feats = feats.mean(-1)
+ feats = feats.view(1, -1)
+ padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+ inputs = {
+ "source": feats,
+ "padding_mask": padding_mask,
+ "output_layer": 9 if self.version == "v1" else 12,
+ }
+ logits = self.hubert.extract_features(**inputs)
+ feats = (
+ self.hubert.final_proj(logits[0]) if self.version == "v1" else logits[0]
+ )
+ feats = torch.cat((feats, feats[:, -1:, :]), 1)
+ if protect < 0.5 and self.if_f0 == 1:
+ feats0 = feats.clone()
+
+ try:
+ if hasattr(self, "index") and self.index_rate > 0:
+ npy = feats[0][skip_head // 2 :].cpu().numpy()
+ if self.is_half:
+ npy = npy.astype("float32")
+ score, ix = self.index.search(npy, k=8)
+ if (ix >= 0).all():
+ weight = np.square(1 / score)
+ weight /= weight.sum(axis=1, keepdims=True)
+ npy = np.sum(
+ self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1
+ )
+ if self.is_half:
+ npy = npy.astype("float16")
+ feats[0][skip_head // 2 :] = (
+ torch.from_numpy(npy).unsqueeze(0).to(self.device)
+ * self.index_rate
+ + (1 - self.index_rate) * feats[0][skip_head // 2 :]
+ )
+ except:
+ pass
+
+ p_len = input_wav.shape[0] // self.window
+ factor = pow(2, self.formant_shift / 12)
+ return_length2 = int(np.ceil(return_length * factor))
+ cache_pitch = cache_pitchf = None
+ pitch = pitchf = None
+ if isinstance(f0method, tuple):
+ pitch, pitchf = f0method
+ pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+ pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+ elif self.if_f0 == 1:
+ f0_extractor_frame = block_frame_16k + 800
+ if f0method == "rmvpe":
+ f0_extractor_frame = (
+ 5120 * ((f0_extractor_frame - 1) // 5120 + 1) - self.window
+ )
+ if inp_f0 is not None:
+ pitch, pitchf = self._get_f0_post(
+ inp_f0, self.f0_up_key - self.formant_shift
+ )
+ else:
+ pitch, pitchf = self._get_f0(
+ input_wav[-f0_extractor_frame:],
+ self.f0_up_key - self.formant_shift,
+ method=f0method,
+ )
+ shift = block_frame_16k // self.window
+ self.cache_pitch[:-shift] = self.cache_pitch[shift:].clone()
+ self.cache_pitchf[:-shift] = self.cache_pitchf[shift:].clone()
+ self.cache_pitch[4 - pitch.shape[0] :] = pitch[3:-1]
+ self.cache_pitchf[4 - pitch.shape[0] :] = pitchf[3:-1]
+ cache_pitch = self.cache_pitch[None, -p_len:]
+ cache_pitchf = (
+ self.cache_pitchf[None, -p_len:] * return_length2 / return_length
+ )
+
+ feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+ feats = feats[:, :p_len, :]
+ if protect < 0.5 and pitch is not None and pitchf is not None:
+ feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
+ 0, 2, 1
+ )
+ feats0 = feats0[:, :p_len, :]
+ pitchff = pitchf.clone()
+ pitchff[pitchf > 0] = 1
+ pitchff[pitchf < 1] = protect
+ pitchff = pitchff.unsqueeze(-1)
+ feats = feats * pitchff + feats0 * (1 - pitchff)
+ feats = feats.to(feats0.dtype)
+ p_len = torch.LongTensor([p_len]).to(self.device)
+ sid = torch.LongTensor([0]).to(self.device)
+ with torch.no_grad():
+ infered_audio = (
+ self.net_g.infer(
+ feats,
+ p_len,
+ sid,
+ pitch=cache_pitch,
+ pitchf=cache_pitchf,
+ skip_head=skip_head,
+ return_length=return_length,
+ return_length2=return_length2,
+ )
+ .squeeze(1)
+ .float()
+ )
+ upp_res = int(np.floor(factor * self.tgt_sr // 100))
+ if upp_res != self.tgt_sr // 100:
+ if upp_res not in self.resample_kernel:
+ self.resample_kernel[upp_res] = Resample(
+ orig_freq=upp_res,
+ new_freq=self.tgt_sr // 100,
+ dtype=torch.float32,
+ ).to(self.device)
+ infered_audio = self.resample_kernel[upp_res](
+ infered_audio[:, : return_length * upp_res]
+ )
+ return infered_audio.squeeze()
+
+ def _get_f0(
+ self,
+ x: torch.Tensor,
+ f0_up_key: Union[int, float],
+ filter_radius: Optional[Union[int, float]] = None,
+ method: Literal["crepe", "rmvpe", "fcpe", "pm", "harvest", "dio"] = "fcpe",
+ ):
+ if method not in self.f0_methods.keys():
+ raise RuntimeError("Not supported f0 method: " + method)
+ return self.f0_methods[method](x, f0_up_key, filter_radius)
+
+ def _get_f0_post(self, f0, f0_up_key):
+ f0 *= pow(2, f0_up_key / 12)
+ if not torch.is_tensor(f0):
+ f0 = torch.from_numpy(f0)
+ f0 = f0.float().to(self.device).squeeze()
+ f0_mel = 1127 * torch.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * 254 / (
+ self.f0_mel_max - self.f0_mel_min
+ ) + 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > 255] = 255
+ f0_coarse = torch.round(f0_mel).long()
+ return f0_coarse, f0
+
+ def _get_f0_pm(self, x, f0_up_key, filter_radius):
+ if not hasattr(self, "pm"):
+ self.pm = PM(hop_length=160, sampling_rate=16000)
+ f0 = self.pm.compute_f0(x.cpu().numpy())
+ return self._get_f0_post(f0, f0_up_key)
+
+ def _get_f0_harvest(self, x, f0_up_key, filter_radius=3):
+ if not hasattr(self, "harvest"):
+ self.harvest = Harvest(
+ self.window,
+ self.f0_min,
+ self.f0_max,
+ self.sr,
+ )
+ if filter_radius is None:
+ filter_radius = 3
+ f0 = self.harvest.compute_f0(x.cpu().numpy(), filter_radius=filter_radius)
+ return self._get_f0_post(f0, f0_up_key)
+
+ def _get_f0_dio(self, x, f0_up_key, filter_radius):
+ if not hasattr(self, "dio"):
+ self.dio = Dio(
+ self.window,
+ self.f0_min,
+ self.f0_max,
+ self.sr,
+ )
+ f0 = self.dio.compute_f0(x.cpu().numpy())
+ return self._get_f0_post(f0, f0_up_key)
+
+ def _get_f0_crepe(self, x, f0_up_key, filter_radius):
+ if hasattr(self, "crepe") == False:
+ self.crepe = CRePE(
+ self.window,
+ self.f0_min,
+ self.f0_max,
+ self.sr,
+ self.device,
+ )
+ f0 = self.crepe.compute_f0(x)
+ return self._get_f0_post(f0, f0_up_key)
+
+ def _get_f0_rmvpe(self, x, f0_up_key, filter_radius=0.03):
+ if hasattr(self, "rmvpe") == False:
+ self.rmvpe = RMVPE(
+ "%s/rmvpe.pt" % os.environ["rmvpe_root"],
+ is_half=self.is_half,
+ device=self.device,
+ use_jit=self.use_jit,
+ )
+ if filter_radius is None:
+ filter_radius = 0.03
+ return self._get_f0_post(
+ self.rmvpe.compute_f0(x, filter_radius=filter_radius),
+ f0_up_key,
+ )
+
+ def _get_f0_fcpe(self, x, f0_up_key, filter_radius):
+ if hasattr(self, "fcpe") == False:
+ self.fcpe = FCPE(
+ 160,
+ self.f0_min,
+ self.f0_max,
+ 16000,
+ self.device,
+ )
+ f0 = self.fcpe.compute_f0(x)
+ return self._get_f0_post(f0, f0_up_key)
diff --git a/infer/lib/rvcmd.py b/infer/lib/rvcmd.py
new file mode 100644
index 0000000000000000000000000000000000000000..d5f6bbd7aa0cd013b6305357d958a31f7681d796
--- /dev/null
+++ b/infer/lib/rvcmd.py
@@ -0,0 +1,241 @@
+import os
+from pathlib import Path
+import hashlib
+import requests
+from io import BytesIO
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def sha256(f) -> str:
+ sha256_hash = hashlib.sha256()
+ # Read and update hash in chunks of 4M
+ for byte_block in iter(lambda: f.read(4 * 1024 * 1024), b""):
+ sha256_hash.update(byte_block)
+ return sha256_hash.hexdigest()
+
+
+def check_model(
+ dir_name: Path, model_name: str, hash: str, remove_incorrect=False
+) -> bool:
+ target = dir_name / model_name
+ relname = target.as_posix()
+ relname = relname[relname.rindex("assets/") :]
+ logger.debug(f"checking {relname}...")
+ if not os.path.exists(target):
+ logger.info(f"{target} not exist.")
+ return False
+ with open(target, "rb") as f:
+ digest = sha256(f)
+ bakfile = f"{target}.bak"
+ if digest != hash:
+ logger.warn(f"{target} sha256 hash mismatch.")
+ logger.info(f"expected: {hash}")
+ logger.info(f"real val: {digest}")
+ logger.warn("please add parameter --update to download the latest assets.")
+ if remove_incorrect:
+ if not os.path.exists(bakfile):
+ os.rename(str(target), bakfile)
+ else:
+ os.remove(str(target))
+ return False
+ if remove_incorrect and os.path.exists(bakfile):
+ os.remove(bakfile)
+ return True
+
+
+def check_all_assets(update=False) -> bool:
+ BASE_DIR = Path(__file__).resolve().parent.parent.parent
+
+ logger.info("checking hubret & rmvpe...")
+
+ if not check_model(
+ BASE_DIR / "assets" / "hubert",
+ "hubert_base.pt",
+ os.environ["sha256_hubert_base_pt"],
+ update,
+ ):
+ return False
+ if not check_model(
+ BASE_DIR / "assets" / "rmvpe",
+ "rmvpe.pt",
+ os.environ["sha256_rmvpe_pt"],
+ update,
+ ):
+ return False
+ if not check_model(
+ BASE_DIR / "assets" / "rmvpe",
+ "rmvpe.onnx",
+ os.environ["sha256_rmvpe_onnx"],
+ update,
+ ):
+ return False
+
+ rvc_models_dir = BASE_DIR / "assets" / "pretrained"
+ logger.info("checking pretrained models...")
+ model_names = [
+ "D32k.pth",
+ "D40k.pth",
+ "D48k.pth",
+ "G32k.pth",
+ "G40k.pth",
+ "G48k.pth",
+ "f0D32k.pth",
+ "f0D40k.pth",
+ "f0D48k.pth",
+ "f0G32k.pth",
+ "f0G40k.pth",
+ "f0G48k.pth",
+ ]
+ for model in model_names:
+ menv = model.replace(".", "_")
+ if not check_model(
+ rvc_models_dir, model, os.environ[f"sha256_v1_{menv}"], update
+ ):
+ return False
+
+ rvc_models_dir = BASE_DIR / "assets" / "pretrained_v2"
+ logger.info("checking pretrained models v2...")
+ for model in model_names:
+ menv = model.replace(".", "_")
+ if not check_model(
+ rvc_models_dir, model, os.environ[f"sha256_v2_{menv}"], update
+ ):
+ return False
+
+ logger.info("checking uvr5_weights...")
+ rvc_models_dir = BASE_DIR / "assets" / "uvr5_weights"
+ model_names = [
+ "HP2-人声vocals+非人声instrumentals.pth",
+ "HP2_all_vocals.pth",
+ "HP3_all_vocals.pth",
+ "HP5-主旋律人声vocals+其他instrumentals.pth",
+ "HP5_only_main_vocal.pth",
+ "VR-DeEchoAggressive.pth",
+ "VR-DeEchoDeReverb.pth",
+ "VR-DeEchoNormal.pth",
+ ]
+ for model in model_names:
+ menv = model.replace(".", "_")
+ if not check_model(
+ rvc_models_dir, model, os.environ[f"sha256_uvr5_{menv}"], update
+ ):
+ return False
+ if not check_model(
+ BASE_DIR / "assets" / "uvr5_weights" / "onnx_dereverb_By_FoxJoy",
+ "vocals.onnx",
+ os.environ[f"sha256_uvr5_vocals_onnx"],
+ update,
+ ):
+ return False
+
+ logger.info("all assets are already latest.")
+ return True
+
+
+def download_and_extract_tar_gz(url: str, folder: str):
+ import tarfile
+
+ logger.info(f"downloading {url}")
+ response = requests.get(url, stream=True, timeout=(5, 10))
+ with BytesIO() as out_file:
+ out_file.write(response.content)
+ out_file.seek(0)
+ logger.info(f"downloaded.")
+ with tarfile.open(fileobj=out_file, mode="r:gz") as tar:
+ tar.extractall(folder)
+ logger.info(f"extracted into {folder}")
+
+
+def download_and_extract_zip(url: str, folder: str):
+ import zipfile
+
+ logger.info(f"downloading {url}")
+ response = requests.get(url, stream=True, timeout=(5, 10))
+ with BytesIO() as out_file:
+ out_file.write(response.content)
+ out_file.seek(0)
+ logger.info(f"downloaded.")
+ with zipfile.ZipFile(out_file) as zip_ref:
+ zip_ref.extractall(folder)
+ logger.info(f"extracted into {folder}")
+
+
+def download_dns_yaml(url: str, folder: str):
+ logger.info(f"downloading {url}")
+ response = requests.get(url, stream=True, timeout=(5, 10))
+ with open(os.path.join(folder, "dns.yaml"), "wb") as out_file:
+ out_file.write(response.content)
+ logger.info(f"downloaded into {folder}")
+
+
+def download_all_assets(tmpdir: str, version="0.2.3"):
+ import subprocess
+ import platform
+
+ archs = {
+ "aarch64": "arm64",
+ "armv8l": "arm64",
+ "arm64": "arm64",
+ "x86": "386",
+ "i386": "386",
+ "i686": "386",
+ "386": "386",
+ "x86_64": "amd64",
+ "x64": "amd64",
+ "amd64": "amd64",
+ }
+ system_type = platform.system().lower()
+ architecture = platform.machine().lower()
+ is_win = system_type == "windows"
+
+ architecture = archs.get(architecture, None)
+ if not architecture:
+ logger.error(f"architecture {architecture} is not supported")
+ exit(1)
+ try:
+ BASE_URL = "https://github.com/fumiama/RVC-Models-Downloader/releases/download/"
+ suffix = "zip" if is_win else "tar.gz"
+ RVCMD_URL = BASE_URL + f"v{version}/rvcmd_{system_type}_{architecture}.{suffix}"
+ cmdfile = os.path.join(tmpdir, "rvcmd")
+ if is_win:
+ download_and_extract_zip(RVCMD_URL, tmpdir)
+ cmdfile += ".exe"
+ else:
+ download_and_extract_tar_gz(RVCMD_URL, tmpdir)
+ os.chmod(cmdfile, 0o755)
+ subprocess.run([cmdfile, "-notui", "-w", "0", "assets/all"])
+ except Exception:
+ BASE_URL = "https://raw.gitcode.com/u011570312/RVC-Models-Downloader/assets/"
+ suffix = {
+ "darwin_amd64": "460",
+ "darwin_arm64": "461",
+ "linux_386": "462",
+ "linux_amd64": "463",
+ "linux_arm64": "464",
+ "windows_386": "465",
+ "windows_amd64": "466",
+ }[f"{system_type}_{architecture}"]
+ RVCMD_URL = BASE_URL + suffix
+ download_dns_yaml(
+ "https://raw.gitcode.com/u011570312/RVC-Models-Downloader/raw/main/dns.yaml",
+ tmpdir,
+ )
+ if is_win:
+ download_and_extract_zip(RVCMD_URL, tmpdir)
+ cmdfile += ".exe"
+ else:
+ download_and_extract_tar_gz(RVCMD_URL, tmpdir)
+ os.chmod(cmdfile, 0o755)
+ subprocess.run(
+ [
+ cmdfile,
+ "-notui",
+ "-w",
+ "0",
+ "-dns",
+ os.path.join(tmpdir, "dns.yaml"),
+ "assets/all",
+ ]
+ )
diff --git a/infer/lib/slicer2.py b/infer/lib/slicer2.py
new file mode 100644
index 0000000000000000000000000000000000000000..7d9d16db55e30c5c732f7fd32a234af026097e13
--- /dev/null
+++ b/infer/lib/slicer2.py
@@ -0,0 +1,260 @@
+import numpy as np
+
+
+# This function is obtained from librosa.
+def get_rms(
+ y,
+ frame_length=2048,
+ hop_length=512,
+ pad_mode="constant",
+):
+ padding = (int(frame_length // 2), int(frame_length // 2))
+ y = np.pad(y, padding, mode=pad_mode)
+
+ axis = -1
+ # put our new within-frame axis at the end for now
+ out_strides = y.strides + tuple([y.strides[axis]])
+ # Reduce the shape on the framing axis
+ x_shape_trimmed = list(y.shape)
+ x_shape_trimmed[axis] -= frame_length - 1
+ out_shape = tuple(x_shape_trimmed) + tuple([frame_length])
+ xw = np.lib.stride_tricks.as_strided(y, shape=out_shape, strides=out_strides)
+ if axis < 0:
+ target_axis = axis - 1
+ else:
+ target_axis = axis + 1
+ xw = np.moveaxis(xw, -1, target_axis)
+ # Downsample along the target axis
+ slices = [slice(None)] * xw.ndim
+ slices[axis] = slice(0, None, hop_length)
+ x = xw[tuple(slices)]
+
+ # Calculate power
+ power = np.mean(np.abs(x) ** 2, axis=-2, keepdims=True)
+
+ return np.sqrt(power)
+
+
+class Slicer:
+ def __init__(
+ self,
+ sr: int,
+ threshold: float = -40.0,
+ min_length: int = 5000,
+ min_interval: int = 300,
+ hop_size: int = 20,
+ max_sil_kept: int = 5000,
+ ):
+ if not min_length >= min_interval >= hop_size:
+ raise ValueError(
+ "The following condition must be satisfied: min_length >= min_interval >= hop_size"
+ )
+ if not max_sil_kept >= hop_size:
+ raise ValueError(
+ "The following condition must be satisfied: max_sil_kept >= hop_size"
+ )
+ min_interval = sr * min_interval / 1000
+ self.threshold = 10 ** (threshold / 20.0)
+ self.hop_size = round(sr * hop_size / 1000)
+ self.win_size = min(round(min_interval), 4 * self.hop_size)
+ self.min_length = round(sr * min_length / 1000 / self.hop_size)
+ self.min_interval = round(min_interval / self.hop_size)
+ self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+ def _apply_slice(self, waveform, begin, end):
+ if len(waveform.shape) > 1:
+ return waveform[
+ :, begin * self.hop_size : min(waveform.shape[1], end * self.hop_size)
+ ]
+ else:
+ return waveform[
+ begin * self.hop_size : min(waveform.shape[0], end * self.hop_size)
+ ]
+
+ # @timeit
+ def slice(self, waveform):
+ if len(waveform.shape) > 1:
+ samples = waveform.mean(axis=0)
+ else:
+ samples = waveform
+ if samples.shape[0] <= self.min_length:
+ return [waveform]
+ rms_list = get_rms(
+ y=samples, frame_length=self.win_size, hop_length=self.hop_size
+ ).squeeze(0)
+ sil_tags = []
+ silence_start = None
+ clip_start = 0
+ for i, rms in enumerate(rms_list):
+ # Keep looping while frame is silent.
+ if rms < self.threshold:
+ # Record start of silent frames.
+ if silence_start is None:
+ silence_start = i
+ continue
+ # Keep looping while frame is not silent and silence start has not been recorded.
+ if silence_start is None:
+ continue
+ # Clear recorded silence start if interval is not enough or clip is too short
+ is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+ need_slice_middle = (
+ i - silence_start >= self.min_interval
+ and i - clip_start >= self.min_length
+ )
+ if not is_leading_silence and not need_slice_middle:
+ silence_start = None
+ continue
+ # Need slicing. Record the range of silent frames to be removed.
+ if i - silence_start <= self.max_sil_kept:
+ pos = rms_list[silence_start : i + 1].argmin() + silence_start
+ if silence_start == 0:
+ sil_tags.append((0, pos))
+ else:
+ sil_tags.append((pos, pos))
+ clip_start = pos
+ elif i - silence_start <= self.max_sil_kept * 2:
+ pos = rms_list[
+ i - self.max_sil_kept : silence_start + self.max_sil_kept + 1
+ ].argmin()
+ pos += i - self.max_sil_kept
+ pos_l = (
+ rms_list[
+ silence_start : silence_start + self.max_sil_kept + 1
+ ].argmin()
+ + silence_start
+ )
+ pos_r = (
+ rms_list[i - self.max_sil_kept : i + 1].argmin()
+ + i
+ - self.max_sil_kept
+ )
+ if silence_start == 0:
+ sil_tags.append((0, pos_r))
+ clip_start = pos_r
+ else:
+ sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+ clip_start = max(pos_r, pos)
+ else:
+ pos_l = (
+ rms_list[
+ silence_start : silence_start + self.max_sil_kept + 1
+ ].argmin()
+ + silence_start
+ )
+ pos_r = (
+ rms_list[i - self.max_sil_kept : i + 1].argmin()
+ + i
+ - self.max_sil_kept
+ )
+ if silence_start == 0:
+ sil_tags.append((0, pos_r))
+ else:
+ sil_tags.append((pos_l, pos_r))
+ clip_start = pos_r
+ silence_start = None
+ # Deal with trailing silence.
+ total_frames = rms_list.shape[0]
+ if (
+ silence_start is not None
+ and total_frames - silence_start >= self.min_interval
+ ):
+ silence_end = min(total_frames, silence_start + self.max_sil_kept)
+ pos = rms_list[silence_start : silence_end + 1].argmin() + silence_start
+ sil_tags.append((pos, total_frames + 1))
+ # Apply and return slices.
+ if len(sil_tags) == 0:
+ return [waveform]
+ else:
+ chunks = []
+ if sil_tags[0][0] > 0:
+ chunks.append(self._apply_slice(waveform, 0, sil_tags[0][0]))
+ for i in range(len(sil_tags) - 1):
+ chunks.append(
+ self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0])
+ )
+ if sil_tags[-1][1] < total_frames:
+ chunks.append(
+ self._apply_slice(waveform, sil_tags[-1][1], total_frames)
+ )
+ return chunks
+
+
+def main():
+ import os.path
+ from argparse import ArgumentParser
+
+ import librosa
+ import soundfile
+
+ parser = ArgumentParser()
+ parser.add_argument("audio", type=str, help="The audio to be sliced")
+ parser.add_argument(
+ "--out", type=str, help="Output directory of the sliced audio clips"
+ )
+ parser.add_argument(
+ "--db_thresh",
+ type=float,
+ required=False,
+ default=-40,
+ help="The dB threshold for silence detection",
+ )
+ parser.add_argument(
+ "--min_length",
+ type=int,
+ required=False,
+ default=5000,
+ help="The minimum milliseconds required for each sliced audio clip",
+ )
+ parser.add_argument(
+ "--min_interval",
+ type=int,
+ required=False,
+ default=300,
+ help="The minimum milliseconds for a silence part to be sliced",
+ )
+ parser.add_argument(
+ "--hop_size",
+ type=int,
+ required=False,
+ default=10,
+ help="Frame length in milliseconds",
+ )
+ parser.add_argument(
+ "--max_sil_kept",
+ type=int,
+ required=False,
+ default=500,
+ help="The maximum silence length kept around the sliced clip, presented in milliseconds",
+ )
+ args = parser.parse_args()
+ out = args.out
+ if out is None:
+ out = os.path.dirname(os.path.abspath(args.audio))
+ audio, sr = librosa.load(args.audio, sr=None, mono=False)
+ slicer = Slicer(
+ sr=sr,
+ threshold=args.db_thresh,
+ min_length=args.min_length,
+ min_interval=args.min_interval,
+ hop_size=args.hop_size,
+ max_sil_kept=args.max_sil_kept,
+ )
+ chunks = slicer.slice(audio)
+ if not os.path.exists(out):
+ os.makedirs(out)
+ for i, chunk in enumerate(chunks):
+ if len(chunk.shape) > 1:
+ chunk = chunk.T
+ soundfile.write(
+ os.path.join(
+ out,
+ f"%s_%d.wav"
+ % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
+ ),
+ chunk,
+ sr,
+ )
+
+
+if __name__ == "__main__":
+ main()
diff --git a/infer/lib/train/data_utils.py b/infer/lib/train/data_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..1e1d1dbfbcba8a83e019950492d047263f7b4e73
--- /dev/null
+++ b/infer/lib/train/data_utils.py
@@ -0,0 +1,517 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import torch
+import torch.utils.data
+
+from infer.lib.train.mel_processing import spectrogram_torch
+from infer.lib.train.utils import load_filepaths_and_text, load_wav_to_torch
+
+
+class TextAudioLoaderMultiNSFsid(torch.utils.data.Dataset):
+ """
+ 1) loads audio, text pairs
+ 2) normalizes text and converts them to sequences of integers
+ 3) computes spectrograms from audio files.
+ """
+
+ def __init__(self, audiopaths_and_text, hparams):
+ self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+ self.max_wav_value = hparams.max_wav_value
+ self.sampling_rate = hparams.sampling_rate
+ self.filter_length = hparams.filter_length
+ self.hop_length = hparams.hop_length
+ self.win_length = hparams.win_length
+ self.sampling_rate = hparams.sampling_rate
+ self.min_text_len = getattr(hparams, "min_text_len", 1)
+ self.max_text_len = getattr(hparams, "max_text_len", 5000)
+ self._filter()
+
+ def _filter(self):
+ """
+ Filter text & store spec lengths
+ """
+ # Store spectrogram lengths for Bucketing
+ # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+ # spec_length = wav_length // hop_length
+ audiopaths_and_text_new = []
+ lengths = []
+ for audiopath, text, pitch, pitchf, dv in self.audiopaths_and_text:
+ if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+ audiopaths_and_text_new.append([audiopath, text, pitch, pitchf, dv])
+ lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+ self.audiopaths_and_text = audiopaths_and_text_new
+ self.lengths = lengths
+
+ def get_sid(self, sid):
+ sid = torch.LongTensor([int(sid)])
+ return sid
+
+ def get_audio_text_pair(self, audiopath_and_text):
+ # separate filename and text
+ file = audiopath_and_text[0]
+ phone = audiopath_and_text[1]
+ pitch = audiopath_and_text[2]
+ pitchf = audiopath_and_text[3]
+ dv = audiopath_and_text[4]
+
+ phone, pitch, pitchf = self.get_labels(phone, pitch, pitchf)
+ spec, wav = self.get_audio(file)
+ dv = self.get_sid(dv)
+
+ len_phone = phone.size()[0]
+ len_spec = spec.size()[-1]
+ # print(123,phone.shape,pitch.shape,spec.shape)
+ if len_phone != len_spec:
+ len_min = min(len_phone, len_spec)
+ # amor
+ len_wav = len_min * self.hop_length
+
+ spec = spec[:, :len_min]
+ wav = wav[:, :len_wav]
+
+ phone = phone[:len_min, :]
+ pitch = pitch[:len_min]
+ pitchf = pitchf[:len_min]
+
+ return (spec, wav, phone, pitch, pitchf, dv)
+
+ def get_labels(self, phone, pitch, pitchf):
+ phone = np.load(phone)
+ phone = np.repeat(phone, 2, axis=0)
+ pitch = np.load(pitch)
+ pitchf = np.load(pitchf)
+ n_num = min(phone.shape[0], 900) # DistributedBucketSampler
+ # print(234,phone.shape,pitch.shape)
+ phone = phone[:n_num, :]
+ pitch = pitch[:n_num]
+ pitchf = pitchf[:n_num]
+ phone = torch.FloatTensor(phone)
+ pitch = torch.LongTensor(pitch)
+ pitchf = torch.FloatTensor(pitchf)
+ return phone, pitch, pitchf
+
+ def get_audio(self, filename):
+ audio, sampling_rate = load_wav_to_torch(filename)
+ if sampling_rate != self.sampling_rate:
+ raise ValueError(
+ "{} SR doesn't match target {} SR".format(
+ sampling_rate, self.sampling_rate
+ )
+ )
+ audio_norm = audio
+ # audio_norm = audio / self.max_wav_value
+ # audio_norm = audio / np.abs(audio).max()
+
+ audio_norm = audio_norm.unsqueeze(0)
+ spec_filename = filename.replace(".wav", ".spec.pt")
+ if os.path.exists(spec_filename):
+ try:
+ spec = torch.load(spec_filename)
+ except:
+ logger.warning("%s %s", spec_filename, traceback.format_exc())
+ spec = spectrogram_torch(
+ audio_norm,
+ self.filter_length,
+ self.sampling_rate,
+ self.hop_length,
+ self.win_length,
+ center=False,
+ )
+ spec = torch.squeeze(spec, 0)
+ torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+ else:
+ spec = spectrogram_torch(
+ audio_norm,
+ self.filter_length,
+ self.sampling_rate,
+ self.hop_length,
+ self.win_length,
+ center=False,
+ )
+ spec = torch.squeeze(spec, 0)
+ torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+ return spec, audio_norm
+
+ def __getitem__(self, index):
+ return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+ def __len__(self):
+ return len(self.audiopaths_and_text)
+
+
+class TextAudioCollateMultiNSFsid:
+ """Zero-pads model inputs and targets"""
+
+ def __init__(self, return_ids=False):
+ self.return_ids = return_ids
+
+ def __call__(self, batch):
+ """Collate's training batch from normalized text and aduio
+ PARAMS
+ ------
+ batch: [text_normalized, spec_normalized, wav_normalized]
+ """
+ # Right zero-pad all one-hot text sequences to max input length
+ _, ids_sorted_decreasing = torch.sort(
+ torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+ )
+
+ max_spec_len = max([x[0].size(1) for x in batch])
+ max_wave_len = max([x[1].size(1) for x in batch])
+ spec_lengths = torch.LongTensor(len(batch))
+ wave_lengths = torch.LongTensor(len(batch))
+ spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+ wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+ spec_padded.zero_()
+ wave_padded.zero_()
+
+ max_phone_len = max([x[2].size(0) for x in batch])
+ phone_lengths = torch.LongTensor(len(batch))
+ phone_padded = torch.FloatTensor(
+ len(batch), max_phone_len, batch[0][2].shape[1]
+ ) # (spec, wav, phone, pitch)
+ pitch_padded = torch.LongTensor(len(batch), max_phone_len)
+ pitchf_padded = torch.FloatTensor(len(batch), max_phone_len)
+ phone_padded.zero_()
+ pitch_padded.zero_()
+ pitchf_padded.zero_()
+ # dv = torch.FloatTensor(len(batch), 256)#gin=256
+ sid = torch.LongTensor(len(batch))
+
+ for i in range(len(ids_sorted_decreasing)):
+ row = batch[ids_sorted_decreasing[i]]
+
+ spec = row[0]
+ spec_padded[i, :, : spec.size(1)] = spec
+ spec_lengths[i] = spec.size(1)
+
+ wave = row[1]
+ wave_padded[i, :, : wave.size(1)] = wave
+ wave_lengths[i] = wave.size(1)
+
+ phone = row[2]
+ phone_padded[i, : phone.size(0), :] = phone
+ phone_lengths[i] = phone.size(0)
+
+ pitch = row[3]
+ pitch_padded[i, : pitch.size(0)] = pitch
+ pitchf = row[4]
+ pitchf_padded[i, : pitchf.size(0)] = pitchf
+
+ # dv[i] = row[5]
+ sid[i] = row[5]
+
+ return (
+ phone_padded,
+ phone_lengths,
+ pitch_padded,
+ pitchf_padded,
+ spec_padded,
+ spec_lengths,
+ wave_padded,
+ wave_lengths,
+ # dv
+ sid,
+ )
+
+
+class TextAudioLoader(torch.utils.data.Dataset):
+ """
+ 1) loads audio, text pairs
+ 2) normalizes text and converts them to sequences of integers
+ 3) computes spectrograms from audio files.
+ """
+
+ def __init__(self, audiopaths_and_text, hparams):
+ self.audiopaths_and_text = load_filepaths_and_text(audiopaths_and_text)
+ self.max_wav_value = hparams.max_wav_value
+ self.sampling_rate = hparams.sampling_rate
+ self.filter_length = hparams.filter_length
+ self.hop_length = hparams.hop_length
+ self.win_length = hparams.win_length
+ self.sampling_rate = hparams.sampling_rate
+ self.min_text_len = getattr(hparams, "min_text_len", 1)
+ self.max_text_len = getattr(hparams, "max_text_len", 5000)
+ self._filter()
+
+ def _filter(self):
+ """
+ Filter text & store spec lengths
+ """
+ # Store spectrogram lengths for Bucketing
+ # wav_length ~= file_size / (wav_channels * Bytes per dim) = file_size / (1 * 2)
+ # spec_length = wav_length // hop_length
+ audiopaths_and_text_new = []
+ lengths = []
+ for audiopath, text, dv in self.audiopaths_and_text:
+ if self.min_text_len <= len(text) and len(text) <= self.max_text_len:
+ audiopaths_and_text_new.append([audiopath, text, dv])
+ lengths.append(os.path.getsize(audiopath) // (3 * self.hop_length))
+ self.audiopaths_and_text = audiopaths_and_text_new
+ self.lengths = lengths
+
+ def get_sid(self, sid):
+ sid = torch.LongTensor([int(sid)])
+ return sid
+
+ def get_audio_text_pair(self, audiopath_and_text):
+ # separate filename and text
+ file = audiopath_and_text[0]
+ phone = audiopath_and_text[1]
+ dv = audiopath_and_text[2]
+
+ phone = self.get_labels(phone)
+ spec, wav = self.get_audio(file)
+ dv = self.get_sid(dv)
+
+ len_phone = phone.size()[0]
+ len_spec = spec.size()[-1]
+ if len_phone != len_spec:
+ len_min = min(len_phone, len_spec)
+ len_wav = len_min * self.hop_length
+ spec = spec[:, :len_min]
+ wav = wav[:, :len_wav]
+ phone = phone[:len_min, :]
+ return (spec, wav, phone, dv)
+
+ def get_labels(self, phone):
+ phone = np.load(phone)
+ phone = np.repeat(phone, 2, axis=0)
+ n_num = min(phone.shape[0], 900) # DistributedBucketSampler
+ phone = phone[:n_num, :]
+ phone = torch.FloatTensor(phone)
+ return phone
+
+ def get_audio(self, filename):
+ audio, sampling_rate = load_wav_to_torch(filename)
+ if sampling_rate != self.sampling_rate:
+ raise ValueError(
+ "{} SR doesn't match target {} SR".format(
+ sampling_rate, self.sampling_rate
+ )
+ )
+ audio_norm = audio
+ # audio_norm = audio / self.max_wav_value
+ # audio_norm = audio / np.abs(audio).max()
+
+ audio_norm = audio_norm.unsqueeze(0)
+ spec_filename = filename.replace(".wav", ".spec.pt")
+ if os.path.exists(spec_filename):
+ try:
+ spec = torch.load(spec_filename)
+ except:
+ logger.warning("%s %s", spec_filename, traceback.format_exc())
+ spec = spectrogram_torch(
+ audio_norm,
+ self.filter_length,
+ self.sampling_rate,
+ self.hop_length,
+ self.win_length,
+ center=False,
+ )
+ spec = torch.squeeze(spec, 0)
+ torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+ else:
+ spec = spectrogram_torch(
+ audio_norm,
+ self.filter_length,
+ self.sampling_rate,
+ self.hop_length,
+ self.win_length,
+ center=False,
+ )
+ spec = torch.squeeze(spec, 0)
+ torch.save(spec, spec_filename, _use_new_zipfile_serialization=False)
+ return spec, audio_norm
+
+ def __getitem__(self, index):
+ return self.get_audio_text_pair(self.audiopaths_and_text[index])
+
+ def __len__(self):
+ return len(self.audiopaths_and_text)
+
+
+class TextAudioCollate:
+ """Zero-pads model inputs and targets"""
+
+ def __init__(self, return_ids=False):
+ self.return_ids = return_ids
+
+ def __call__(self, batch):
+ """Collate's training batch from normalized text and aduio
+ PARAMS
+ ------
+ batch: [text_normalized, spec_normalized, wav_normalized]
+ """
+ # Right zero-pad all one-hot text sequences to max input length
+ _, ids_sorted_decreasing = torch.sort(
+ torch.LongTensor([x[0].size(1) for x in batch]), dim=0, descending=True
+ )
+
+ max_spec_len = max([x[0].size(1) for x in batch])
+ max_wave_len = max([x[1].size(1) for x in batch])
+ spec_lengths = torch.LongTensor(len(batch))
+ wave_lengths = torch.LongTensor(len(batch))
+ spec_padded = torch.FloatTensor(len(batch), batch[0][0].size(0), max_spec_len)
+ wave_padded = torch.FloatTensor(len(batch), 1, max_wave_len)
+ spec_padded.zero_()
+ wave_padded.zero_()
+
+ max_phone_len = max([x[2].size(0) for x in batch])
+ phone_lengths = torch.LongTensor(len(batch))
+ phone_padded = torch.FloatTensor(
+ len(batch), max_phone_len, batch[0][2].shape[1]
+ )
+ phone_padded.zero_()
+ sid = torch.LongTensor(len(batch))
+
+ for i in range(len(ids_sorted_decreasing)):
+ row = batch[ids_sorted_decreasing[i]]
+
+ spec = row[0]
+ spec_padded[i, :, : spec.size(1)] = spec
+ spec_lengths[i] = spec.size(1)
+
+ wave = row[1]
+ wave_padded[i, :, : wave.size(1)] = wave
+ wave_lengths[i] = wave.size(1)
+
+ phone = row[2]
+ phone_padded[i, : phone.size(0), :] = phone
+ phone_lengths[i] = phone.size(0)
+
+ sid[i] = row[3]
+
+ return (
+ phone_padded,
+ phone_lengths,
+ spec_padded,
+ spec_lengths,
+ wave_padded,
+ wave_lengths,
+ sid,
+ )
+
+
+class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
+ """
+ Maintain similar input lengths in a batch.
+ Length groups are specified by boundaries.
+ Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
+
+ It removes samples which are not included in the boundaries.
+ Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
+ """
+
+ def __init__(
+ self,
+ dataset,
+ batch_size,
+ boundaries,
+ num_replicas=None,
+ rank=None,
+ shuffle=True,
+ ):
+ super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
+ self.lengths = dataset.lengths
+ self.batch_size = batch_size
+ self.boundaries = boundaries
+
+ self.buckets, self.num_samples_per_bucket = self._create_buckets()
+ self.total_size = sum(self.num_samples_per_bucket)
+ self.num_samples = self.total_size // self.num_replicas
+
+ def _create_buckets(self):
+ buckets = [[] for _ in range(len(self.boundaries) - 1)]
+ for i in range(len(self.lengths)):
+ length = self.lengths[i]
+ idx_bucket = self._bisect(length)
+ if idx_bucket != -1:
+ buckets[idx_bucket].append(i)
+
+ for i in range(len(buckets) - 1, -1, -1): #
+ if len(buckets[i]) == 0:
+ buckets.pop(i)
+ self.boundaries.pop(i + 1)
+
+ num_samples_per_bucket = []
+ for i in range(len(buckets)):
+ len_bucket = len(buckets[i])
+ total_batch_size = self.num_replicas * self.batch_size
+ rem = (
+ total_batch_size - (len_bucket % total_batch_size)
+ ) % total_batch_size
+ num_samples_per_bucket.append(len_bucket + rem)
+ return buckets, num_samples_per_bucket
+
+ def __iter__(self):
+ # deterministically shuffle based on epoch
+ g = torch.Generator()
+ g.manual_seed(self.epoch)
+
+ indices = []
+ if self.shuffle:
+ for bucket in self.buckets:
+ indices.append(torch.randperm(len(bucket), generator=g).tolist())
+ else:
+ for bucket in self.buckets:
+ indices.append(list(range(len(bucket))))
+
+ batches = []
+ for i in range(len(self.buckets)):
+ bucket = self.buckets[i]
+ len_bucket = len(bucket)
+ ids_bucket = indices[i]
+ num_samples_bucket = self.num_samples_per_bucket[i]
+
+ # add extra samples to make it evenly divisible
+ rem = num_samples_bucket - len_bucket
+ ids_bucket = (
+ ids_bucket
+ + ids_bucket * (rem // len_bucket)
+ + ids_bucket[: (rem % len_bucket)]
+ )
+
+ # subsample
+ ids_bucket = ids_bucket[self.rank :: self.num_replicas]
+
+ # batching
+ for j in range(len(ids_bucket) // self.batch_size):
+ batch = [
+ bucket[idx]
+ for idx in ids_bucket[
+ j * self.batch_size : (j + 1) * self.batch_size
+ ]
+ ]
+ batches.append(batch)
+
+ if self.shuffle:
+ batch_ids = torch.randperm(len(batches), generator=g).tolist()
+ batches = [batches[i] for i in batch_ids]
+ self.batches = batches
+
+ assert len(self.batches) * self.batch_size == self.num_samples
+ return iter(self.batches)
+
+ def _bisect(self, x, lo=0, hi=None):
+ if hi is None:
+ hi = len(self.boundaries) - 1
+
+ if hi > lo:
+ mid = (hi + lo) // 2
+ if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
+ return mid
+ elif x <= self.boundaries[mid]:
+ return self._bisect(x, lo, mid)
+ else:
+ return self._bisect(x, mid + 1, hi)
+ else:
+ return -1
+
+ def __len__(self):
+ return self.num_samples // self.batch_size
diff --git a/infer/lib/train/losses.py b/infer/lib/train/losses.py
new file mode 100644
index 0000000000000000000000000000000000000000..aa7bd81cf596884a8b33e802ae49254d7810a860
--- /dev/null
+++ b/infer/lib/train/losses.py
@@ -0,0 +1,58 @@
+import torch
+
+
+def feature_loss(fmap_r, fmap_g):
+ loss = 0
+ for dr, dg in zip(fmap_r, fmap_g):
+ for rl, gl in zip(dr, dg):
+ rl = rl.float().detach()
+ gl = gl.float()
+ loss += torch.mean(torch.abs(rl - gl))
+
+ return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+ loss = 0
+ r_losses = []
+ g_losses = []
+ for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+ dr = dr.float()
+ dg = dg.float()
+ r_loss = torch.mean((1 - dr) ** 2)
+ g_loss = torch.mean(dg**2)
+ loss += r_loss + g_loss
+ r_losses.append(r_loss.item())
+ g_losses.append(g_loss.item())
+
+ return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+ loss = 0
+ gen_losses = []
+ for dg in disc_outputs:
+ dg = dg.float()
+ l = torch.mean((1 - dg) ** 2)
+ gen_losses.append(l)
+ loss += l
+
+ return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+ """
+ z_p, logs_q: [b, h, t_t]
+ m_p, logs_p: [b, h, t_t]
+ """
+ z_p = z_p.float()
+ logs_q = logs_q.float()
+ m_p = m_p.float()
+ logs_p = logs_p.float()
+ z_mask = z_mask.float()
+
+ kl = logs_p - logs_q - 0.5
+ kl += 0.5 * ((z_p - m_p) ** 2) * torch.exp(-2.0 * logs_p)
+ kl = torch.sum(kl * z_mask)
+ l = kl / torch.sum(z_mask)
+ return l
diff --git a/infer/lib/train/mel_processing.py b/infer/lib/train/mel_processing.py
new file mode 100644
index 0000000000000000000000000000000000000000..3751f1eab1ea8137088f2f7d7c8294190403b4ce
--- /dev/null
+++ b/infer/lib/train/mel_processing.py
@@ -0,0 +1,127 @@
+import torch
+import torch.utils.data
+from librosa.filters import mel as librosa_mel_fn
+import logging
+
+logger = logging.getLogger(__name__)
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+ """
+ PARAMS
+ ------
+ C: compression factor
+ """
+ return torch.log(torch.clamp(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+ """
+ PARAMS
+ ------
+ C: compression factor used to compress
+ """
+ return torch.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+ return dynamic_range_compression_torch(magnitudes)
+
+
+def spectral_de_normalize_torch(magnitudes):
+ return dynamic_range_decompression_torch(magnitudes)
+
+
+# Reusable banks
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+ """Convert waveform into Linear-frequency Linear-amplitude spectrogram.
+
+ Args:
+ y :: (B, T) - Audio waveforms
+ n_fft
+ sampling_rate
+ hop_size
+ win_size
+ center
+ Returns:
+ :: (B, Freq, Frame) - Linear-frequency Linear-amplitude spectrogram
+ """
+
+ # Window - Cache if needed
+ global hann_window
+ dtype_device = str(y.dtype) + "_" + str(y.device)
+ wnsize_dtype_device = str(win_size) + "_" + dtype_device
+ if wnsize_dtype_device not in hann_window:
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(
+ dtype=y.dtype, device=y.device
+ )
+
+ # Padding
+ y = torch.nn.functional.pad(
+ y.unsqueeze(1),
+ (int((n_fft - hop_size) / 2), int((n_fft - hop_size) / 2)),
+ mode="reflect",
+ )
+ y = y.squeeze(1)
+
+ # Complex Spectrogram :: (B, T) -> (B, Freq, Frame, RealComplex=2)
+ spec = torch.stft(
+ y,
+ n_fft,
+ hop_length=hop_size,
+ win_length=win_size,
+ window=hann_window[wnsize_dtype_device],
+ center=center,
+ pad_mode="reflect",
+ normalized=False,
+ onesided=True,
+ return_complex=True,
+ )
+
+ # Linear-frequency Linear-amplitude spectrogram :: (B, Freq, Frame, RealComplex=2) -> (B, Freq, Frame)
+ spec = torch.sqrt(spec.real.pow(2) + spec.imag.pow(2) + 1e-6)
+ return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+ # MelBasis - Cache if needed
+ global mel_basis
+ dtype_device = str(spec.dtype) + "_" + str(spec.device)
+ fmax_dtype_device = str(fmax) + "_" + dtype_device
+ if fmax_dtype_device not in mel_basis:
+ mel = librosa_mel_fn(
+ sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax
+ )
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(
+ dtype=spec.dtype, device=spec.device
+ )
+
+ # Mel-frequency Log-amplitude spectrogram :: (B, Freq=num_mels, Frame)
+ melspec = torch.matmul(mel_basis[fmax_dtype_device], spec)
+ melspec = spectral_normalize_torch(melspec)
+ return melspec
+
+
+def mel_spectrogram_torch(
+ y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False
+):
+ """Convert waveform into Mel-frequency Log-amplitude spectrogram.
+
+ Args:
+ y :: (B, T) - Waveforms
+ Returns:
+ melspec :: (B, Freq, Frame) - Mel-frequency Log-amplitude spectrogram
+ """
+ # Linear-frequency Linear-amplitude spectrogram :: (B, T) -> (B, Freq, Frame)
+ spec = spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center)
+
+ # Mel-frequency Log-amplitude spectrogram :: (B, Freq, Frame) -> (B, Freq=num_mels, Frame)
+ melspec = spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax)
+
+ return melspec
diff --git a/infer/lib/train/process_ckpt.py b/infer/lib/train/process_ckpt.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf9697fe7389f32879f30db56e3ad99c07fe417a
--- /dev/null
+++ b/infer/lib/train/process_ckpt.py
@@ -0,0 +1,280 @@
+import os
+import traceback
+from collections import OrderedDict
+from time import time
+
+import torch
+
+from i18n.i18n import I18nAuto
+from infer.modules.vc import model_hash_ckpt, hash_id
+
+i18n = I18nAuto()
+
+
+# add author sign
+def save_small_model(ckpt, sr, if_f0, name, epoch, version, hps):
+ try:
+ opt = OrderedDict()
+ opt["weight"] = {}
+ for key in ckpt.keys():
+ if "enc_q" in key:
+ continue
+ opt["weight"][key] = ckpt[key].half()
+ opt["config"] = [
+ hps.data.filter_length // 2 + 1,
+ 32,
+ hps.model.inter_channels,
+ hps.model.hidden_channels,
+ hps.model.filter_channels,
+ hps.model.n_heads,
+ hps.model.n_layers,
+ hps.model.kernel_size,
+ hps.model.p_dropout,
+ hps.model.resblock,
+ hps.model.resblock_kernel_sizes,
+ hps.model.resblock_dilation_sizes,
+ hps.model.upsample_rates,
+ hps.model.upsample_initial_channel,
+ hps.model.upsample_kernel_sizes,
+ hps.model.spk_embed_dim,
+ hps.model.gin_channels,
+ hps.data.sampling_rate,
+ ]
+ opt["info"] = "%sepoch" % epoch
+ opt["name"] = name
+ opt["timestamp"] = int(time())
+ if hps.author:
+ opt["author"] = hps.author
+ opt["sr"] = sr
+ opt["f0"] = if_f0
+ opt["version"] = version
+ h = model_hash_ckpt(opt)
+ opt["hash"] = h
+ opt["id"] = hash_id(h)
+ torch.save(opt, "assets/weights/%s.pth" % name)
+ return "Success."
+ except:
+ return traceback.format_exc()
+
+
+def extract_small_model(path, name, author, sr, if_f0, info, version):
+ try:
+ ckpt = torch.load(path, map_location="cpu")
+ if "model" in ckpt:
+ ckpt = ckpt["model"]
+ opt = OrderedDict()
+ opt["weight"] = {}
+ for key in ckpt.keys():
+ if "enc_q" in key:
+ continue
+ opt["weight"][key] = ckpt[key].half()
+ if sr == "40k":
+ opt["config"] = [
+ 1025,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [10, 10, 2, 2],
+ 512,
+ [16, 16, 4, 4],
+ 109,
+ 256,
+ 40000,
+ ]
+ elif sr == "48k":
+ if version == "v1":
+ opt["config"] = [
+ 1025,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [10, 6, 2, 2, 2],
+ 512,
+ [16, 16, 4, 4, 4],
+ 109,
+ 256,
+ 48000,
+ ]
+ else:
+ opt["config"] = [
+ 1025,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [12, 10, 2, 2],
+ 512,
+ [24, 20, 4, 4],
+ 109,
+ 256,
+ 48000,
+ ]
+ elif sr == "32k":
+ if version == "v1":
+ opt["config"] = [
+ 513,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [10, 4, 2, 2, 2],
+ 512,
+ [16, 16, 4, 4, 4],
+ 109,
+ 256,
+ 32000,
+ ]
+ else:
+ opt["config"] = [
+ 513,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [10, 8, 2, 2],
+ 512,
+ [20, 16, 4, 4],
+ 109,
+ 256,
+ 32000,
+ ]
+ if info == "":
+ info = "Extracted model."
+ opt["info"] = info
+ opt["name"] = name
+ opt["timestamp"] = int(time())
+ if author:
+ opt["author"] = author
+ opt["version"] = version
+ opt["sr"] = sr
+ opt["f0"] = int(if_f0)
+ h = model_hash_ckpt(opt)
+ opt["hash"] = h
+ opt["id"] = hash_id(h)
+ torch.save(opt, "assets/weights/%s.pth" % name)
+ return "Success."
+ except:
+ return traceback.format_exc()
+
+
+def change_info(path, info, name):
+ try:
+ ckpt = torch.load(path, map_location="cpu")
+ ckpt["info"] = info
+ if name == "":
+ name = os.path.basename(path)
+ torch.save(ckpt, "assets/weights/%s" % name)
+ return "Success."
+ except:
+ return traceback.format_exc()
+
+
+def merge(path1, path2, alpha1, sr, f0, info, name, version):
+ try:
+
+ def extract(ckpt):
+ a = ckpt["model"]
+ opt = OrderedDict()
+ opt["weight"] = {}
+ for key in a.keys():
+ if "enc_q" in key:
+ continue
+ opt["weight"][key] = a[key]
+ return opt
+
+ def authors(c1, c2):
+ a1, a2 = c1.get("author", ""), c2.get("author", "")
+ if a1 == a2:
+ return a1
+ if not a1:
+ a1 = "Unknown"
+ if not a2:
+ a2 = "Unknown"
+ return f"{a1} & {a2}"
+
+ ckpt1 = torch.load(path1, map_location="cpu")
+ ckpt2 = torch.load(path2, map_location="cpu")
+ cfg = ckpt1["config"]
+ if "model" in ckpt1:
+ ckpt1 = extract(ckpt1)
+ else:
+ ckpt1 = ckpt1["weight"]
+ if "model" in ckpt2:
+ ckpt2 = extract(ckpt2)
+ else:
+ ckpt2 = ckpt2["weight"]
+ if sorted(list(ckpt1.keys())) != sorted(list(ckpt2.keys())):
+ return "Fail to merge the models. The model architectures are not the same."
+ opt = OrderedDict()
+ opt["weight"] = {}
+ for key in ckpt1.keys():
+ # try:
+ if key == "emb_g.weight" and ckpt1[key].shape != ckpt2[key].shape:
+ min_shape0 = min(ckpt1[key].shape[0], ckpt2[key].shape[0])
+ opt["weight"][key] = (
+ alpha1 * (ckpt1[key][:min_shape0].float())
+ + (1 - alpha1) * (ckpt2[key][:min_shape0].float())
+ ).half()
+ else:
+ opt["weight"][key] = (
+ alpha1 * (ckpt1[key].float()) + (1 - alpha1) * (ckpt2[key].float())
+ ).half()
+ author = authors(ckpt1, ckpt2)
+ opt["config"] = cfg
+ """
+ if(sr=="40k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 10, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 40000]
+ elif(sr=="48k"):opt["config"] = [1025, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,6,2,2,2], 512, [16, 16, 4, 4], 109, 256, 48000]
+ elif(sr=="32k"):opt["config"] = [513, 32, 192, 192, 768, 2, 6, 3, 0, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10, 4, 2, 2, 2], 512, [16, 16, 4, 4,4], 109, 256, 32000]
+ """
+ opt["name"] = name
+ opt["timestamp"] = int(time())
+ if author:
+ opt["author"] = author
+ opt["sr"] = sr
+ opt["f0"] = 1 if f0 == i18n("Yes") else 0
+ opt["version"] = version
+ opt["info"] = info
+ h = model_hash_ckpt(opt)
+ opt["hash"] = h
+ opt["id"] = hash_id(h)
+ torch.save(opt, "assets/weights/%s.pth" % name)
+ return "Success."
+ except:
+ return traceback.format_exc()
diff --git a/infer/lib/train/utils.py b/infer/lib/train/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..bad2c78a86f1bd1c20fcb380987d3db5bde9a460
--- /dev/null
+++ b/infer/lib/train/utils.py
@@ -0,0 +1,465 @@
+import argparse
+import glob
+import json
+import logging
+import os
+import sys
+from copy import deepcopy
+
+import codecs
+import numpy as np
+import torch
+from scipy.io.wavfile import read
+
+MATPLOTLIB_FLAG = False
+
+logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
+logger = logging
+
+"""
+def load_checkpoint_d(checkpoint_path, combd, sbd, optimizer=None, load_opt=1):
+ assert os.path.isfile(checkpoint_path)
+ checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+
+ ##################
+ def go(model, bkey):
+ saved_state_dict = checkpoint_dict[bkey]
+ if hasattr(model, "module"):
+ state_dict = model.module.state_dict()
+ else:
+ state_dict = model.state_dict()
+ new_state_dict = {}
+ for k, v in state_dict.items(): # 模型需要的shape
+ try:
+ new_state_dict[k] = saved_state_dict[k]
+ if saved_state_dict[k].shape != state_dict[k].shape:
+ logger.warning(
+ "shape-%s-mismatch. need: %s, get: %s",
+ k,
+ state_dict[k].shape,
+ saved_state_dict[k].shape,
+ ) #
+ raise KeyError
+ except:
+ # logger.info(traceback.format_exc())
+ logger.info("%s is not in the checkpoint", k) # pretrain缺失的
+ new_state_dict[k] = v # 模型自带的随机值
+ if hasattr(model, "module"):
+ model.module.load_state_dict(new_state_dict, strict=False)
+ else:
+ model.load_state_dict(new_state_dict, strict=False)
+ return model
+
+ go(combd, "combd")
+ model = go(sbd, "sbd")
+ #############
+ logger.info("Loaded model weights")
+
+ iteration = checkpoint_dict["iteration"]
+ learning_rate = checkpoint_dict["learning_rate"]
+ if (
+ optimizer is not None and load_opt == 1
+ ): ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+ # try:
+ optimizer.load_state_dict(checkpoint_dict["optimizer"])
+ # except:
+ # traceback.print_exc()
+ logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+ return model, optimizer, learning_rate, iteration
+"""
+
+
+def load_checkpoint(checkpoint_path, model, optimizer=None, load_opt=1):
+ assert os.path.isfile(checkpoint_path)
+ checkpoint_dict = torch.load(checkpoint_path, map_location="cpu")
+
+ saved_state_dict = checkpoint_dict["model"]
+ if hasattr(model, "module"):
+ state_dict = model.module.state_dict()
+ else:
+ state_dict = model.state_dict()
+ new_state_dict = {}
+ for k, v in state_dict.items(): # 模型需要的shape
+ try:
+ new_state_dict[k] = saved_state_dict[k]
+ if saved_state_dict[k].shape != state_dict[k].shape:
+ logger.warning(
+ "shape-%s-mismatch|need-%s|get-%s",
+ k,
+ state_dict[k].shape,
+ saved_state_dict[k].shape,
+ ) #
+ raise KeyError
+ except:
+ # logger.info(traceback.format_exc())
+ logger.info("%s is not in the checkpoint", k) # pretrain缺失的
+ new_state_dict[k] = v # 模型自带的随机值
+ if hasattr(model, "module"):
+ model.module.load_state_dict(new_state_dict, strict=False)
+ else:
+ model.load_state_dict(new_state_dict, strict=False)
+ logger.info("Loaded model weights")
+
+ iteration = checkpoint_dict["iteration"]
+ learning_rate = checkpoint_dict["learning_rate"]
+ if (
+ optimizer is not None and load_opt == 1
+ ): ###加载不了，如果是空的的话，重新初始化，可能还会影响lr时间表的更新，因此在train文件最外围catch
+ # try:
+ optimizer.load_state_dict(checkpoint_dict["optimizer"])
+ # except:
+ # traceback.print_exc()
+ logger.info("Loaded checkpoint '{}' (epoch {})".format(checkpoint_path, iteration))
+ return model, optimizer, learning_rate, iteration
+
+
+def save_checkpoint(model, optimizer, learning_rate, iteration, checkpoint_path):
+ logger.info(
+ "Saving model and optimizer state at epoch {} to {}".format(
+ iteration, checkpoint_path
+ )
+ )
+ if hasattr(model, "module"):
+ state_dict = model.module.state_dict()
+ else:
+ state_dict = model.state_dict()
+ torch.save(
+ {
+ "model": state_dict,
+ "iteration": iteration,
+ "optimizer": optimizer.state_dict(),
+ "learning_rate": learning_rate,
+ },
+ checkpoint_path,
+ )
+
+
+"""
+def save_checkpoint_d(combd, sbd, optimizer, learning_rate, iteration, checkpoint_path):
+ logger.info(
+ "Saving model and optimizer state at epoch {} to {}".format(
+ iteration, checkpoint_path
+ )
+ )
+ if hasattr(combd, "module"):
+ state_dict_combd = combd.module.state_dict()
+ else:
+ state_dict_combd = combd.state_dict()
+ if hasattr(sbd, "module"):
+ state_dict_sbd = sbd.module.state_dict()
+ else:
+ state_dict_sbd = sbd.state_dict()
+ torch.save(
+ {
+ "combd": state_dict_combd,
+ "sbd": state_dict_sbd,
+ "iteration": iteration,
+ "optimizer": optimizer.state_dict(),
+ "learning_rate": learning_rate,
+ },
+ checkpoint_path,
+ )
+"""
+
+
+def summarize(
+ writer,
+ global_step,
+ scalars={},
+ histograms={},
+ images={},
+ audios={},
+ audio_sampling_rate=22050,
+):
+ for k, v in scalars.items():
+ writer.add_scalar(k, v, global_step)
+ for k, v in histograms.items():
+ writer.add_histogram(k, v, global_step)
+ for k, v in images.items():
+ writer.add_image(k, v, global_step, dataformats="HWC")
+ for k, v in audios.items():
+ writer.add_audio(k, v, global_step, audio_sampling_rate)
+
+
+def latest_checkpoint_path(dir_path, regex="G_*.pth"):
+ f_list = glob.glob(os.path.join(dir_path, regex))
+ f_list.sort(key=lambda f: int("".join(filter(str.isdigit, f))))
+ x = f_list[-1]
+ logger.debug(x)
+ return x
+
+
+def plot_spectrogram_to_numpy(spectrogram):
+ global MATPLOTLIB_FLAG
+ if not MATPLOTLIB_FLAG:
+ import matplotlib
+
+ matplotlib.use("Agg")
+ MATPLOTLIB_FLAG = True
+ mpl_logger = logging.getLogger("matplotlib")
+ mpl_logger.setLevel(logging.WARNING)
+ import matplotlib.pylab as plt
+ import numpy as np
+
+ fig, ax = plt.subplots(figsize=(10, 2))
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower", interpolation="none")
+ plt.colorbar(im, ax=ax)
+ plt.xlabel("Frames")
+ plt.ylabel("Channels")
+ plt.tight_layout()
+
+ fig.canvas.draw()
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+ plt.close()
+ return data
+
+
+def plot_alignment_to_numpy(alignment, info=None):
+ global MATPLOTLIB_FLAG
+ if not MATPLOTLIB_FLAG:
+ import matplotlib
+
+ matplotlib.use("Agg")
+ MATPLOTLIB_FLAG = True
+ mpl_logger = logging.getLogger("matplotlib")
+ mpl_logger.setLevel(logging.WARNING)
+ import matplotlib.pylab as plt
+ import numpy as np
+
+ fig, ax = plt.subplots(figsize=(6, 4))
+ im = ax.imshow(
+ alignment.transpose(), aspect="auto", origin="lower", interpolation="none"
+ )
+ fig.colorbar(im, ax=ax)
+ xlabel = "Decoder timestep"
+ if info is not None:
+ xlabel += "\n\n" + info
+ plt.xlabel(xlabel)
+ plt.ylabel("Encoder timestep")
+ plt.tight_layout()
+
+ fig.canvas.draw()
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep="")
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
+ plt.close()
+ return data
+
+
+def load_wav_to_torch(full_path):
+ sampling_rate, data = read(full_path)
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
+
+
+def load_filepaths_and_text(filename, split="|"):
+ try:
+ return [
+ line.strip().split(split)
+ for line in codecs.open(filename, encoding="utf-8")
+ ]
+ except UnicodeDecodeError as e:
+ logger.error("Error loading file %s: %s", filename, e)
+
+ return []
+
+
+def get_hparams(init=True):
+ """
+ todo:
+ 结尾七人组：
+ 保存频率、总epoch done
+ bs done
+ pretrainG、pretrainD done
+ 卡号：os.en["CUDA_VISIBLE_DEVICES"] done
+ if_latest done
+ 模型：if_f0 done
+ 采样率：自动选择config done
+ 是否缓存数据集进GPU:if_cache_data_in_gpu done
+
+ -m:
+ 自动决定training_files路径,改掉train_nsf_load_pretrain.py里的hps.data.training_files done
+ -c不要了
+ """
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ "-se",
+ "--save_every_epoch",
+ type=int,
+ required=True,
+ help="checkpoint save frequency (epoch)",
+ )
+ parser.add_argument(
+ "-te", "--total_epoch", type=int, required=True, help="total_epoch"
+ )
+ parser.add_argument(
+ "-pg", "--pretrainG", type=str, default="", help="Pretrained Generator path"
+ )
+ parser.add_argument(
+ "-pd", "--pretrainD", type=str, default="", help="Pretrained Discriminator path"
+ )
+ parser.add_argument("-g", "--gpus", type=str, default="0", help="split by -")
+ parser.add_argument(
+ "-bs", "--batch_size", type=int, required=True, help="batch size"
+ )
+ parser.add_argument(
+ "-e", "--experiment_dir", type=str, required=True, help="experiment dir"
+ ) # -m
+ parser.add_argument(
+ "-sr", "--sample_rate", type=str, required=True, help="sample rate, 32k/40k/48k"
+ )
+ parser.add_argument(
+ "-sw",
+ "--save_every_weights",
+ type=str,
+ default="0",
+ help="save the extracted model in weights directory when saving checkpoints",
+ )
+ parser.add_argument(
+ "-v", "--version", type=str, required=True, help="model version"
+ )
+ parser.add_argument(
+ "-f0",
+ "--if_f0",
+ type=int,
+ required=True,
+ help="use f0 as one of the inputs of the model, 1 or 0",
+ )
+ parser.add_argument(
+ "-l",
+ "--if_latest",
+ type=int,
+ required=True,
+ help="if only save the latest G/D pth file, 1 or 0",
+ )
+ parser.add_argument(
+ "-c",
+ "--if_cache_data_in_gpu",
+ type=int,
+ required=True,
+ help="if caching the dataset in GPU memory, 1 or 0",
+ )
+ parser.add_argument("-a", "--author", type=str, default="", help="Model author")
+
+ args = parser.parse_args()
+ name = args.experiment_dir
+ experiment_dir = os.path.join("./logs", args.experiment_dir)
+
+ config_save_path = os.path.join(experiment_dir, "config.json")
+ with open(config_save_path, "r") as f:
+ config = json.load(f)
+
+ hparams = HParams(**config)
+ hparams.model_dir = hparams.experiment_dir = experiment_dir
+ hparams.save_every_epoch = args.save_every_epoch
+ hparams.name = name
+ hparams.total_epoch = args.total_epoch
+ hparams.pretrainG = args.pretrainG
+ hparams.pretrainD = args.pretrainD
+ hparams.version = args.version
+ hparams.gpus = args.gpus
+ hparams.train.batch_size = args.batch_size
+ hparams.sample_rate = args.sample_rate
+ hparams.if_f0 = args.if_f0
+ hparams.if_latest = args.if_latest
+ hparams.save_every_weights = args.save_every_weights
+ hparams.if_cache_data_in_gpu = args.if_cache_data_in_gpu
+ hparams.data.training_files = "%s/filelist.txt" % experiment_dir
+ hparams.author = args.author
+ return hparams
+
+
+"""
+def get_hparams_from_dir(model_dir):
+ config_save_path = os.path.join(model_dir, "config.json")
+ with open(config_save_path, "r") as f:
+ data = f.read()
+ config = json.loads(data)
+
+ hparams = HParams(**config)
+ hparams.model_dir = model_dir
+ return hparams
+
+
+def get_hparams_from_file(config_path):
+ with open(config_path, "r") as f:
+ data = f.read()
+ config = json.loads(data)
+
+ hparams = HParams(**config)
+ return hparams
+
+
+def check_git_hash(model_dir):
+ source_dir = os.path.dirname(os.path.realpath(__file__))
+ if not os.path.exists(os.path.join(source_dir, ".git")):
+ logger.warning(
+ "{} is not a git repository, therefore hash value comparison will be ignored.".format(
+ source_dir
+ )
+ )
+ return
+
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
+
+ path = os.path.join(model_dir, "githash")
+ if os.path.exists(path):
+ saved_hash = open(path).read()
+ if saved_hash != cur_hash:
+ logger.warning(
+ "git hash values are different. {}(saved) != {}(current)".format(
+ saved_hash[:8], cur_hash[:8]
+ )
+ )
+ else:
+ open(path, "w").write(cur_hash)
+"""
+
+
+def get_logger(model_dir, filename="train.log"):
+ global logger
+ logger = logging.getLogger(os.path.basename(model_dir))
+ logger.setLevel(logging.DEBUG)
+
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
+ if not os.path.exists(model_dir):
+ os.makedirs(model_dir)
+ h = logging.FileHandler(os.path.join(model_dir, filename))
+ h.setLevel(logging.DEBUG)
+ h.setFormatter(formatter)
+ logger.addHandler(h)
+ return logger
+
+
+class HParams:
+ def __init__(self, **kwargs):
+ for k, v in kwargs.items():
+ if type(v) == dict:
+ v = HParams(**v)
+ self[k] = v
+
+ def keys(self):
+ return self.__dict__.keys()
+
+ def items(self):
+ return self.__dict__.items()
+
+ def values(self):
+ return self.__dict__.values()
+
+ def copy(self):
+ return deepcopy(self)
+
+ def __len__(self):
+ return len(self.__dict__)
+
+ def __getitem__(self, key):
+ return getattr(self, key)
+
+ def __setitem__(self, key, value):
+ return setattr(self, key, value)
+
+ def __contains__(self, key):
+ return key in self.__dict__
+
+ def __repr__(self):
+ return self.__dict__.__repr__()
diff --git a/infer/lib/uvr5_pack/lib_v5/dataset.py b/infer/lib/uvr5_pack/lib_v5/dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..cfd01a174978d97180a897e40cb59ecadec1d12e
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/dataset.py
@@ -0,0 +1,183 @@
+import os
+import random
+
+import numpy as np
+import torch
+import torch.utils.data
+from tqdm import tqdm
+
+from . import spec_utils
+
+
+class VocalRemoverValidationSet(torch.utils.data.Dataset):
+ def __init__(self, patch_list):
+ self.patch_list = patch_list
+
+ def __len__(self):
+ return len(self.patch_list)
+
+ def __getitem__(self, idx):
+ path = self.patch_list[idx]
+ data = np.load(path)
+
+ X, y = data["X"], data["y"]
+
+ X_mag = np.abs(X)
+ y_mag = np.abs(y)
+
+ return X_mag, y_mag
+
+
+def make_pair(mix_dir, inst_dir):
+ input_exts = [".wav", ".m4a", ".mp3", ".mp4", ".flac"]
+
+ X_list = sorted(
+ [
+ os.path.join(mix_dir, fname)
+ for fname in os.listdir(mix_dir)
+ if os.path.splitext(fname)[1] in input_exts
+ ]
+ )
+ y_list = sorted(
+ [
+ os.path.join(inst_dir, fname)
+ for fname in os.listdir(inst_dir)
+ if os.path.splitext(fname)[1] in input_exts
+ ]
+ )
+
+ filelist = list(zip(X_list, y_list))
+
+ return filelist
+
+
+def train_val_split(dataset_dir, split_mode, val_rate, val_filelist):
+ if split_mode == "random":
+ filelist = make_pair(
+ os.path.join(dataset_dir, "mixtures"),
+ os.path.join(dataset_dir, "instruments"),
+ )
+
+ random.shuffle(filelist)
+
+ if len(val_filelist) == 0:
+ val_size = int(len(filelist) * val_rate)
+ train_filelist = filelist[:-val_size]
+ val_filelist = filelist[-val_size:]
+ else:
+ train_filelist = [
+ pair for pair in filelist if list(pair) not in val_filelist
+ ]
+ elif split_mode == "subdirs":
+ if len(val_filelist) != 0:
+ raise ValueError(
+ "The `val_filelist` option is not available in `subdirs` mode"
+ )
+
+ train_filelist = make_pair(
+ os.path.join(dataset_dir, "training/mixtures"),
+ os.path.join(dataset_dir, "training/instruments"),
+ )
+
+ val_filelist = make_pair(
+ os.path.join(dataset_dir, "validation/mixtures"),
+ os.path.join(dataset_dir, "validation/instruments"),
+ )
+
+ return train_filelist, val_filelist
+
+
+def augment(X, y, reduction_rate, reduction_mask, mixup_rate, mixup_alpha):
+ perm = np.random.permutation(len(X))
+ for i, idx in enumerate(tqdm(perm)):
+ if np.random.uniform() < reduction_rate:
+ y[idx] = spec_utils.reduce_vocal_aggressively(
+ X[idx], y[idx], reduction_mask
+ )
+
+ if np.random.uniform() < 0.5:
+ # swap channel
+ X[idx] = X[idx, ::-1]
+ y[idx] = y[idx, ::-1]
+ if np.random.uniform() < 0.02:
+ # mono
+ X[idx] = X[idx].mean(axis=0, keepdims=True)
+ y[idx] = y[idx].mean(axis=0, keepdims=True)
+ if np.random.uniform() < 0.02:
+ # inst
+ X[idx] = y[idx]
+
+ if np.random.uniform() < mixup_rate and i < len(perm) - 1:
+ lam = np.random.beta(mixup_alpha, mixup_alpha)
+ X[idx] = lam * X[idx] + (1 - lam) * X[perm[i + 1]]
+ y[idx] = lam * y[idx] + (1 - lam) * y[perm[i + 1]]
+
+ return X, y
+
+
+def make_padding(width, cropsize, offset):
+ left = offset
+ roi_size = cropsize - left * 2
+ if roi_size == 0:
+ roi_size = cropsize
+ right = roi_size - (width % roi_size) + left
+
+ return left, right, roi_size
+
+
+def make_training_set(filelist, cropsize, patches, sr, hop_length, n_fft, offset):
+ len_dataset = patches * len(filelist)
+
+ X_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+ y_dataset = np.zeros((len_dataset, 2, n_fft // 2 + 1, cropsize), dtype=np.complex64)
+
+ for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+ X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+ coef = np.max([np.abs(X).max(), np.abs(y).max()])
+ X, y = X / coef, y / coef
+
+ l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+ X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+ y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+ starts = np.random.randint(0, X_pad.shape[2] - cropsize, patches)
+ ends = starts + cropsize
+ for j in range(patches):
+ idx = i * patches + j
+ X_dataset[idx] = X_pad[:, :, starts[j] : ends[j]]
+ y_dataset[idx] = y_pad[:, :, starts[j] : ends[j]]
+
+ return X_dataset, y_dataset
+
+
+def make_validation_set(filelist, cropsize, sr, hop_length, n_fft, offset):
+ patch_list = []
+ patch_dir = "cs{}_sr{}_hl{}_nf{}_of{}".format(
+ cropsize, sr, hop_length, n_fft, offset
+ )
+ os.makedirs(patch_dir, exist_ok=True)
+
+ for i, (X_path, y_path) in enumerate(tqdm(filelist)):
+ basename = os.path.splitext(os.path.basename(X_path))[0]
+
+ X, y = spec_utils.cache_or_load(X_path, y_path, sr, hop_length, n_fft)
+ coef = np.max([np.abs(X).max(), np.abs(y).max()])
+ X, y = X / coef, y / coef
+
+ l, r, roi_size = make_padding(X.shape[2], cropsize, offset)
+ X_pad = np.pad(X, ((0, 0), (0, 0), (l, r)), mode="constant")
+ y_pad = np.pad(y, ((0, 0), (0, 0), (l, r)), mode="constant")
+
+ len_dataset = int(np.ceil(X.shape[2] / roi_size))
+ for j in range(len_dataset):
+ outpath = os.path.join(patch_dir, "{}_p{}.npz".format(basename, j))
+ start = j * roi_size
+ if not os.path.exists(outpath):
+ np.savez(
+ outpath,
+ X=X_pad[:, :, start : start + cropsize],
+ y=y_pad[:, :, start : start + cropsize],
+ )
+ patch_list.append(outpath)
+
+ return VocalRemoverValidationSet(patch_list)
diff --git a/infer/lib/uvr5_pack/lib_v5/layers.py b/infer/lib/uvr5_pack/lib_v5/layers.py
new file mode 100644
index 0000000000000000000000000000000000000000..44153b6a23399c6938affc61c71919eaa172bcee
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/layers.py
@@ -0,0 +1,125 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(Conv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nout,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ bias=False,
+ ),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class Encoder(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+ super(Encoder, self).__init__()
+ self.conv1 = Conv2DBNActiv(nin, nout, ksize, stride, pad, activ=activ)
+ self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+
+ def __call__(self, x):
+ h = self.conv1(x)
+ h = self.conv2(h)
+
+ return h
+
+
+class Decoder(nn.Module):
+ def __init__(
+ self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+ ):
+ super(Decoder, self).__init__()
+ self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ # self.conv2 = Conv2DBNActiv(nout, nout, ksize, 1, pad, activ=activ)
+ self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+ def __call__(self, x, skip=None):
+ x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+
+ if skip is not None:
+ skip = spec_utils.crop_center(skip, x)
+ x = torch.cat([x, skip], dim=1)
+
+ h = self.conv1(x)
+ # h = self.conv2(h)
+
+ if self.dropout is not None:
+ h = self.dropout(h)
+
+ return h
+
+
+class ASPPModule(nn.Module):
+ def __init__(self, nin, nout, dilations=(4, 8, 12), activ=nn.ReLU, dropout=False):
+ super(ASPPModule, self).__init__()
+ self.conv1 = nn.Sequential(
+ nn.AdaptiveAvgPool2d((1, None)),
+ Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ),
+ )
+ self.conv2 = Conv2DBNActiv(nin, nout, 1, 1, 0, activ=activ)
+ self.conv3 = Conv2DBNActiv(
+ nin, nout, 3, 1, dilations[0], dilations[0], activ=activ
+ )
+ self.conv4 = Conv2DBNActiv(
+ nin, nout, 3, 1, dilations[1], dilations[1], activ=activ
+ )
+ self.conv5 = Conv2DBNActiv(
+ nin, nout, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.bottleneck = Conv2DBNActiv(nout * 5, nout, 1, 1, 0, activ=activ)
+ self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+ def forward(self, x):
+ _, _, h, w = x.size()
+ feat1 = F.interpolate(
+ self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+ )
+ feat2 = self.conv2(x)
+ feat3 = self.conv3(x)
+ feat4 = self.conv4(x)
+ feat5 = self.conv5(x)
+ out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+ out = self.bottleneck(out)
+
+ if self.dropout is not None:
+ out = self.dropout(out)
+
+ return out
+
+
+class LSTMModule(nn.Module):
+ def __init__(self, nin_conv, nin_lstm, nout_lstm):
+ super(LSTMModule, self).__init__()
+ self.conv = Conv2DBNActiv(nin_conv, 1, 1, 1, 0)
+ self.lstm = nn.LSTM(
+ input_size=nin_lstm, hidden_size=nout_lstm // 2, bidirectional=True
+ )
+ self.dense = nn.Sequential(
+ nn.Linear(nout_lstm, nin_lstm), nn.BatchNorm1d(nin_lstm), nn.ReLU()
+ )
+
+ def forward(self, x):
+ N, _, nbins, nframes = x.size()
+ h = self.conv(x)[:, 0] # N, nbins, nframes
+ h = h.permute(2, 0, 1) # nframes, N, nbins
+ h, _ = self.lstm(h)
+ h = self.dense(h.reshape(-1, h.size()[-1])) # nframes * N, nbins
+ h = h.reshape(nframes, N, 1, nbins)
+ h = h.permute(1, 2, 3, 0)
+
+ return h
diff --git a/infer/lib/uvr5_pack/lib_v5/layers_123821KB.py b/infer/lib/uvr5_pack/lib_v5/layers_123821KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..4fc1b5cb85a3327f60cbb9f5deffbeeaaac516ad
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/layers_123821KB.py
@@ -0,0 +1,118 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(Conv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nout,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ bias=False,
+ ),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(SeperableConv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nin,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ groups=nin,
+ bias=False,
+ ),
+ nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class Encoder(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+ super(Encoder, self).__init__()
+ self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+ def __call__(self, x):
+ skip = self.conv1(x)
+ h = self.conv2(skip)
+
+ return h, skip
+
+
+class Decoder(nn.Module):
+ def __init__(
+ self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+ ):
+ super(Decoder, self).__init__()
+ self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+ def __call__(self, x, skip=None):
+ x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+ if skip is not None:
+ skip = spec_utils.crop_center(skip, x)
+ x = torch.cat([x, skip], dim=1)
+ h = self.conv(x)
+
+ if self.dropout is not None:
+ h = self.dropout(h)
+
+ return h
+
+
+class ASPPModule(nn.Module):
+ def __init__(self, nin, nout, dilations=(4, 8, 16), activ=nn.ReLU):
+ super(ASPPModule, self).__init__()
+ self.conv1 = nn.Sequential(
+ nn.AdaptiveAvgPool2d((1, None)),
+ Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+ )
+ self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+ self.conv3 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+ )
+ self.conv4 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+ )
+ self.conv5 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.bottleneck = nn.Sequential(
+ Conv2DBNActiv(nin * 5, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+ )
+
+ def forward(self, x):
+ _, _, h, w = x.size()
+ feat1 = F.interpolate(
+ self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+ )
+ feat2 = self.conv2(x)
+ feat3 = self.conv3(x)
+ feat4 = self.conv4(x)
+ feat5 = self.conv5(x)
+ out = torch.cat((feat1, feat2, feat3, feat4, feat5), dim=1)
+ bottle = self.bottleneck(out)
+ return bottle
diff --git a/infer/lib/uvr5_pack/lib_v5/layers_33966KB.py b/infer/lib/uvr5_pack/lib_v5/layers_33966KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..9b127bc6427f5c60c8cf85603a3d8a093c3501c4
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/layers_33966KB.py
@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(Conv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nout,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ bias=False,
+ ),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(SeperableConv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nin,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ groups=nin,
+ bias=False,
+ ),
+ nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class Encoder(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+ super(Encoder, self).__init__()
+ self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+ def __call__(self, x):
+ skip = self.conv1(x)
+ h = self.conv2(skip)
+
+ return h, skip
+
+
+class Decoder(nn.Module):
+ def __init__(
+ self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+ ):
+ super(Decoder, self).__init__()
+ self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+ def __call__(self, x, skip=None):
+ x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+ if skip is not None:
+ skip = spec_utils.crop_center(skip, x)
+ x = torch.cat([x, skip], dim=1)
+ h = self.conv(x)
+
+ if self.dropout is not None:
+ h = self.dropout(h)
+
+ return h
+
+
+class ASPPModule(nn.Module):
+ def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+ super(ASPPModule, self).__init__()
+ self.conv1 = nn.Sequential(
+ nn.AdaptiveAvgPool2d((1, None)),
+ Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+ )
+ self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+ self.conv3 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+ )
+ self.conv4 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+ )
+ self.conv5 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.conv6 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.conv7 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.bottleneck = nn.Sequential(
+ Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+ )
+
+ def forward(self, x):
+ _, _, h, w = x.size()
+ feat1 = F.interpolate(
+ self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+ )
+ feat2 = self.conv2(x)
+ feat3 = self.conv3(x)
+ feat4 = self.conv4(x)
+ feat5 = self.conv5(x)
+ feat6 = self.conv6(x)
+ feat7 = self.conv7(x)
+ out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+ bottle = self.bottleneck(out)
+ return bottle
diff --git a/infer/lib/uvr5_pack/lib_v5/layers_537238KB.py b/infer/lib/uvr5_pack/lib_v5/layers_537238KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..9b127bc6427f5c60c8cf85603a3d8a093c3501c4
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/layers_537238KB.py
@@ -0,0 +1,126 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import spec_utils
+
+
+class Conv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(Conv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nout,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ bias=False,
+ ),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class SeperableConv2DBNActiv(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, dilation=1, activ=nn.ReLU):
+ super(SeperableConv2DBNActiv, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ nin,
+ nin,
+ kernel_size=ksize,
+ stride=stride,
+ padding=pad,
+ dilation=dilation,
+ groups=nin,
+ bias=False,
+ ),
+ nn.Conv2d(nin, nout, kernel_size=1, bias=False),
+ nn.BatchNorm2d(nout),
+ activ(),
+ )
+
+ def __call__(self, x):
+ return self.conv(x)
+
+
+class Encoder(nn.Module):
+ def __init__(self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.LeakyReLU):
+ super(Encoder, self).__init__()
+ self.conv1 = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.conv2 = Conv2DBNActiv(nout, nout, ksize, stride, pad, activ=activ)
+
+ def __call__(self, x):
+ skip = self.conv1(x)
+ h = self.conv2(skip)
+
+ return h, skip
+
+
+class Decoder(nn.Module):
+ def __init__(
+ self, nin, nout, ksize=3, stride=1, pad=1, activ=nn.ReLU, dropout=False
+ ):
+ super(Decoder, self).__init__()
+ self.conv = Conv2DBNActiv(nin, nout, ksize, 1, pad, activ=activ)
+ self.dropout = nn.Dropout2d(0.1) if dropout else None
+
+ def __call__(self, x, skip=None):
+ x = F.interpolate(x, scale_factor=2, mode="bilinear", align_corners=True)
+ if skip is not None:
+ skip = spec_utils.crop_center(skip, x)
+ x = torch.cat([x, skip], dim=1)
+ h = self.conv(x)
+
+ if self.dropout is not None:
+ h = self.dropout(h)
+
+ return h
+
+
+class ASPPModule(nn.Module):
+ def __init__(self, nin, nout, dilations=(4, 8, 16, 32, 64), activ=nn.ReLU):
+ super(ASPPModule, self).__init__()
+ self.conv1 = nn.Sequential(
+ nn.AdaptiveAvgPool2d((1, None)),
+ Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ),
+ )
+ self.conv2 = Conv2DBNActiv(nin, nin, 1, 1, 0, activ=activ)
+ self.conv3 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[0], dilations[0], activ=activ
+ )
+ self.conv4 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[1], dilations[1], activ=activ
+ )
+ self.conv5 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.conv6 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.conv7 = SeperableConv2DBNActiv(
+ nin, nin, 3, 1, dilations[2], dilations[2], activ=activ
+ )
+ self.bottleneck = nn.Sequential(
+ Conv2DBNActiv(nin * 7, nout, 1, 1, 0, activ=activ), nn.Dropout2d(0.1)
+ )
+
+ def forward(self, x):
+ _, _, h, w = x.size()
+ feat1 = F.interpolate(
+ self.conv1(x), size=(h, w), mode="bilinear", align_corners=True
+ )
+ feat2 = self.conv2(x)
+ feat3 = self.conv3(x)
+ feat4 = self.conv4(x)
+ feat5 = self.conv5(x)
+ feat6 = self.conv6(x)
+ feat7 = self.conv7(x)
+ out = torch.cat((feat1, feat2, feat3, feat4, feat5, feat6, feat7), dim=1)
+ bottle = self.bottleneck(out)
+ return bottle
diff --git a/infer/lib/uvr5_pack/lib_v5/model_param_init.py b/infer/lib/uvr5_pack/lib_v5/model_param_init.py
new file mode 100644
index 0000000000000000000000000000000000000000..b995c0bfb1194746187692e2ab1c2a6dbaaaec6c
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/model_param_init.py
@@ -0,0 +1,69 @@
+import json
+import os
+import pathlib
+
+default_param = {}
+default_param["bins"] = 768
+default_param["unstable_bins"] = 9 # training only
+default_param["reduction_bins"] = 762 # training only
+default_param["sr"] = 44100
+default_param["pre_filter_start"] = 757
+default_param["pre_filter_stop"] = 768
+default_param["band"] = {}
+
+
+default_param["band"][1] = {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 960,
+ "crop_start": 0,
+ "crop_stop": 245,
+ "lpf_start": 61, # inference only
+ "res_type": "polyphase",
+}
+
+default_param["band"][2] = {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 1536,
+ "crop_start": 24,
+ "crop_stop": 547,
+ "hpf_start": 81, # inference only
+ "res_type": "sinc_best",
+}
+
+
+def int_keys(d):
+ r = {}
+ for k, v in d:
+ if k.isdigit():
+ k = int(k)
+ r[k] = v
+ return r
+
+
+class ModelParameters(object):
+ def __init__(self, config_path=""):
+ if ".pth" == pathlib.Path(config_path).suffix:
+ import zipfile
+
+ with zipfile.ZipFile(config_path, "r") as zip:
+ self.param = json.loads(
+ zip.read("param.json"), object_pairs_hook=int_keys
+ )
+ elif ".json" == pathlib.Path(config_path).suffix:
+ with open(config_path, "r") as f:
+ self.param = json.loads(f.read(), object_pairs_hook=int_keys)
+ else:
+ self.param = default_param
+
+ for k in [
+ "mid_side",
+ "mid_side_b",
+ "mid_side_b2",
+ "stereo_w",
+ "stereo_n",
+ "reverse",
+ ]:
+ if not k in self.param:
+ self.param[k] = False
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json
new file mode 100644
index 0000000000000000000000000000000000000000..72cb4499867ad2827185e85687f06fb73d33eced
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 16000,
+ "hl": 512,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 1024,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 16000,
+ "pre_filter_start": 1023,
+ "pre_filter_stop": 1024
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json
new file mode 100644
index 0000000000000000000000000000000000000000..3c00ecf0a105e55a6a86a3c32db301a2635b5b41
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 32000,
+ "hl": 512,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 1024,
+ "hpf_start": -1,
+ "res_type": "kaiser_fast"
+ }
+ },
+ "sr": 32000,
+ "pre_filter_start": 1000,
+ "pre_filter_stop": 1021
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json
new file mode 100644
index 0000000000000000000000000000000000000000..55666ac9a8d0547751fb4b4d3bffb1ee2c956913
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 33075,
+ "hl": 384,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 1024,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 33075,
+ "pre_filter_start": 1000,
+ "pre_filter_stop": 1021
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json
new file mode 100644
index 0000000000000000000000000000000000000000..665abe20eb3cc39fe0f8493dad8f25f6ef634a14
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 44100,
+ "hl": 1024,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 1024,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 1023,
+ "pre_filter_stop": 1024
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json
new file mode 100644
index 0000000000000000000000000000000000000000..0e8b16f89b0231d06eabe8d2f7c2670c7caa2272
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json
@@ -0,0 +1,19 @@
+{
+ "bins": 256,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 44100,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 0,
+ "crop_stop": 256,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 256,
+ "pre_filter_stop": 256
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json
new file mode 100644
index 0000000000000000000000000000000000000000..3b38fcaf60ba204e03a47f5bd3f5bcfe75e1983a
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 1024,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 1023,
+ "pre_filter_stop": 1024
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512_cut.json b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512_cut.json
new file mode 100644
index 0000000000000000000000000000000000000000..630df3524e340f43a1ddb7b33ff02cc91fc1cb47
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512_cut.json
@@ -0,0 +1,19 @@
+{
+ "bins": 1024,
+ "unstable_bins": 0,
+ "reduction_bins": 0,
+ "band": {
+ "1": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 700,
+ "hpf_start": -1,
+ "res_type": "sinc_best"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 1023,
+ "pre_filter_stop": 700
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json
new file mode 100644
index 0000000000000000000000000000000000000000..ab9cf1150a818eb6252105408311be0a40d423b3
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json
@@ -0,0 +1,30 @@
+{
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 705,
+ "band": {
+ "1": {
+ "sr": 6000,
+ "hl": 66,
+ "n_fft": 512,
+ "crop_start": 0,
+ "crop_stop": 240,
+ "lpf_start": 60,
+ "lpf_stop": 118,
+ "res_type": "sinc_fastest"
+ },
+ "2": {
+ "sr": 32000,
+ "hl": 352,
+ "n_fft": 1024,
+ "crop_start": 22,
+ "crop_stop": 505,
+ "hpf_start": 44,
+ "hpf_stop": 23,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 32000,
+ "pre_filter_start": 710,
+ "pre_filter_stop": 731
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json
new file mode 100644
index 0000000000000000000000000000000000000000..7faa216d7b49aeece24123dbdd868847a1dbc03c
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json
@@ -0,0 +1,30 @@
+{
+ "bins": 512,
+ "unstable_bins": 7,
+ "reduction_bins": 510,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 160,
+ "n_fft": 768,
+ "crop_start": 0,
+ "crop_stop": 192,
+ "lpf_start": 41,
+ "lpf_stop": 139,
+ "res_type": "sinc_fastest"
+ },
+ "2": {
+ "sr": 44100,
+ "hl": 640,
+ "n_fft": 1024,
+ "crop_start": 10,
+ "crop_stop": 320,
+ "hpf_start": 47,
+ "hpf_stop": 15,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 510,
+ "pre_filter_stop": 512
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json
new file mode 100644
index 0000000000000000000000000000000000000000..7e78175052b09cb1a32345e54006475992712f9a
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json
@@ -0,0 +1,30 @@
+{
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 705,
+ "band": {
+ "1": {
+ "sr": 6000,
+ "hl": 66,
+ "n_fft": 512,
+ "crop_start": 0,
+ "crop_stop": 240,
+ "lpf_start": 60,
+ "lpf_stop": 240,
+ "res_type": "sinc_fastest"
+ },
+ "2": {
+ "sr": 48000,
+ "hl": 528,
+ "n_fft": 1536,
+ "crop_start": 22,
+ "crop_stop": 505,
+ "hpf_start": 82,
+ "hpf_stop": 22,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 48000,
+ "pre_filter_start": 710,
+ "pre_filter_stop": 731
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json
new file mode 100644
index 0000000000000000000000000000000000000000..d881d767ff83fbac0e18dfe2587ef16925b29b3c
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json
@@ -0,0 +1,42 @@
+{
+ "bins": 768,
+ "unstable_bins": 5,
+ "reduction_bins": 733,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 768,
+ "crop_start": 0,
+ "crop_stop": 278,
+ "lpf_start": 28,
+ "lpf_stop": 140,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 768,
+ "crop_start": 14,
+ "crop_stop": 322,
+ "hpf_start": 70,
+ "hpf_stop": 14,
+ "lpf_start": 283,
+ "lpf_stop": 314,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 131,
+ "crop_stop": 313,
+ "hpf_start": 154,
+ "hpf_stop": 141,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 757,
+ "pre_filter_stop": 768
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json
new file mode 100644
index 0000000000000000000000000000000000000000..77ec198573b19f36519a028a509767d30764c0e2
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json
@@ -0,0 +1,43 @@
+{
+ "mid_side": true,
+ "bins": 768,
+ "unstable_bins": 5,
+ "reduction_bins": 733,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 768,
+ "crop_start": 0,
+ "crop_stop": 278,
+ "lpf_start": 28,
+ "lpf_stop": 140,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 768,
+ "crop_start": 14,
+ "crop_stop": 322,
+ "hpf_start": 70,
+ "hpf_stop": 14,
+ "lpf_start": 283,
+ "lpf_stop": 314,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 131,
+ "crop_stop": 313,
+ "hpf_start": 154,
+ "hpf_stop": 141,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 757,
+ "pre_filter_stop": 768
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json
new file mode 100644
index 0000000000000000000000000000000000000000..85ee8a7d44541c9176e85ea3dce8728d34990938
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json
@@ -0,0 +1,43 @@
+{
+ "mid_side_b2": true,
+ "bins": 640,
+ "unstable_bins": 7,
+ "reduction_bins": 565,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 108,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 187,
+ "lpf_start": 92,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 22050,
+ "hl": 216,
+ "n_fft": 768,
+ "crop_start": 0,
+ "crop_stop": 212,
+ "hpf_start": 68,
+ "hpf_stop": 34,
+ "lpf_start": 174,
+ "lpf_stop": 209,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 44100,
+ "hl": 432,
+ "n_fft": 640,
+ "crop_start": 66,
+ "crop_stop": 307,
+ "hpf_start": 86,
+ "hpf_stop": 72,
+ "res_type": "kaiser_fast"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 639,
+ "pre_filter_stop": 640
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json
new file mode 100644
index 0000000000000000000000000000000000000000..df123754204372aa50d464fbe9102a401f48cc73
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json
@@ -0,0 +1,54 @@
+{
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json
new file mode 100644
index 0000000000000000000000000000000000000000..e91b699eb63d3382c3b9e9edf46d40ed91d6122b
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json
@@ -0,0 +1,55 @@
+{
+ "bins": 768,
+ "unstable_bins": 7,
+ "mid_side": true,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json
new file mode 100644
index 0000000000000000000000000000000000000000..f852f280ec9d98fc1b65cec688290eaafec61b84
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json
@@ -0,0 +1,55 @@
+{
+ "mid_side_b": true,
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json
new file mode 100644
index 0000000000000000000000000000000000000000..f852f280ec9d98fc1b65cec688290eaafec61b84
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json
@@ -0,0 +1,55 @@
+{
+ "mid_side_b": true,
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a07d5541bd83dc1caa20b531c3b43a2ffccac88
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json
@@ -0,0 +1,55 @@
+{
+ "reverse": true,
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba0cf342106de793e6ec3e876854c7fd451fbf76
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json
@@ -0,0 +1,55 @@
+{
+ "stereo_w": true,
+ "bins": 768,
+ "unstable_bins": 7,
+ "reduction_bins": 668,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 1024,
+ "crop_start": 0,
+ "crop_stop": 186,
+ "lpf_start": 37,
+ "lpf_stop": 73,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 11025,
+ "hl": 128,
+ "n_fft": 512,
+ "crop_start": 4,
+ "crop_stop": 185,
+ "hpf_start": 36,
+ "hpf_stop": 18,
+ "lpf_start": 93,
+ "lpf_stop": 185,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 22050,
+ "hl": 256,
+ "n_fft": 512,
+ "crop_start": 46,
+ "crop_stop": 186,
+ "hpf_start": 93,
+ "hpf_stop": 46,
+ "lpf_start": 164,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 512,
+ "n_fft": 768,
+ "crop_start": 121,
+ "crop_stop": 382,
+ "hpf_start": 138,
+ "hpf_stop": 123,
+ "res_type": "sinc_medium"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 740,
+ "pre_filter_stop": 768
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json
new file mode 100644
index 0000000000000000000000000000000000000000..33281a0cf9916fc33558ddfda7a0287a2547faf4
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json
@@ -0,0 +1,54 @@
+{
+ "bins": 672,
+ "unstable_bins": 8,
+ "reduction_bins": 637,
+ "band": {
+ "1": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 640,
+ "crop_start": 0,
+ "crop_stop": 85,
+ "lpf_start": 25,
+ "lpf_stop": 53,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 320,
+ "crop_start": 4,
+ "crop_stop": 87,
+ "hpf_start": 25,
+ "hpf_stop": 12,
+ "lpf_start": 31,
+ "lpf_stop": 62,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 14700,
+ "hl": 160,
+ "n_fft": 512,
+ "crop_start": 17,
+ "crop_stop": 216,
+ "hpf_start": 48,
+ "hpf_stop": 24,
+ "lpf_start": 139,
+ "lpf_stop": 210,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 480,
+ "n_fft": 960,
+ "crop_start": 78,
+ "crop_stop": 383,
+ "hpf_start": 130,
+ "hpf_stop": 86,
+ "res_type": "kaiser_fast"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 668,
+ "pre_filter_stop": 672
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json
new file mode 100644
index 0000000000000000000000000000000000000000..2e5c770fe188779bf6b0873190b7a324d6a867b2
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json
@@ -0,0 +1,55 @@
+{
+ "bins": 672,
+ "unstable_bins": 8,
+ "reduction_bins": 637,
+ "band": {
+ "1": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 640,
+ "crop_start": 0,
+ "crop_stop": 85,
+ "lpf_start": 25,
+ "lpf_stop": 53,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 320,
+ "crop_start": 4,
+ "crop_stop": 87,
+ "hpf_start": 25,
+ "hpf_stop": 12,
+ "lpf_start": 31,
+ "lpf_stop": 62,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 14700,
+ "hl": 160,
+ "n_fft": 512,
+ "crop_start": 17,
+ "crop_stop": 216,
+ "hpf_start": 48,
+ "hpf_stop": 24,
+ "lpf_start": 139,
+ "lpf_stop": 210,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 480,
+ "n_fft": 960,
+ "crop_start": 78,
+ "crop_stop": 383,
+ "hpf_start": 130,
+ "hpf_stop": 86,
+ "convert_channels": "stereo_n",
+ "res_type": "kaiser_fast"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 668,
+ "pre_filter_stop": 672
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v3.json b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v3.json
new file mode 100644
index 0000000000000000000000000000000000000000..2a73bc97ac545145a75bdca7addc5d59f5b8574b
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/4band_v3.json
@@ -0,0 +1,54 @@
+{
+ "bins": 672,
+ "unstable_bins": 8,
+ "reduction_bins": 530,
+ "band": {
+ "1": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 640,
+ "crop_start": 0,
+ "crop_stop": 85,
+ "lpf_start": 25,
+ "lpf_stop": 53,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 7350,
+ "hl": 80,
+ "n_fft": 320,
+ "crop_start": 4,
+ "crop_stop": 87,
+ "hpf_start": 25,
+ "hpf_stop": 12,
+ "lpf_start": 31,
+ "lpf_stop": 62,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 14700,
+ "hl": 160,
+ "n_fft": 512,
+ "crop_start": 17,
+ "crop_stop": 216,
+ "hpf_start": 48,
+ "hpf_stop": 24,
+ "lpf_start": 139,
+ "lpf_stop": 210,
+ "res_type": "polyphase"
+ },
+ "4": {
+ "sr": 44100,
+ "hl": 480,
+ "n_fft": 960,
+ "crop_start": 78,
+ "crop_stop": 383,
+ "hpf_start": 130,
+ "hpf_stop": 86,
+ "res_type": "kaiser_fast"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 668,
+ "pre_filter_stop": 672
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/modelparams/ensemble.json b/infer/lib/uvr5_pack/lib_v5/modelparams/ensemble.json
new file mode 100644
index 0000000000000000000000000000000000000000..ee69beb46fc82f34619c5e48761e329fcabbbd00
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/modelparams/ensemble.json
@@ -0,0 +1,43 @@
+{
+ "mid_side_b2": true,
+ "bins": 1280,
+ "unstable_bins": 7,
+ "reduction_bins": 565,
+ "band": {
+ "1": {
+ "sr": 11025,
+ "hl": 108,
+ "n_fft": 2048,
+ "crop_start": 0,
+ "crop_stop": 374,
+ "lpf_start": 92,
+ "lpf_stop": 186,
+ "res_type": "polyphase"
+ },
+ "2": {
+ "sr": 22050,
+ "hl": 216,
+ "n_fft": 1536,
+ "crop_start": 0,
+ "crop_stop": 424,
+ "hpf_start": 68,
+ "hpf_stop": 34,
+ "lpf_start": 348,
+ "lpf_stop": 418,
+ "res_type": "polyphase"
+ },
+ "3": {
+ "sr": 44100,
+ "hl": 432,
+ "n_fft": 1280,
+ "crop_start": 132,
+ "crop_stop": 614,
+ "hpf_start": 172,
+ "hpf_stop": 144,
+ "res_type": "polyphase"
+ }
+ },
+ "sr": 44100,
+ "pre_filter_start": 1280,
+ "pre_filter_stop": 1280
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/lib_v5/nets.py b/infer/lib/uvr5_pack/lib_v5/nets.py
new file mode 100644
index 0000000000000000000000000000000000000000..6224533b0d8e2aea8d83feb9cc3bdec6cc80a2a4
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/nets.py
@@ -0,0 +1,133 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers
+
+
+class BaseNet(nn.Module):
+ def __init__(
+ self, nin, nout, nin_lstm, nout_lstm, dilations=((4, 2), (8, 4), (12, 6))
+ ):
+ super(BaseNet, self).__init__()
+ self.enc1 = layers.Conv2DBNActiv(nin, nout, 3, 1, 1)
+ self.enc2 = layers.Encoder(nout, nout * 2, 3, 2, 1)
+ self.enc3 = layers.Encoder(nout * 2, nout * 4, 3, 2, 1)
+ self.enc4 = layers.Encoder(nout * 4, nout * 6, 3, 2, 1)
+ self.enc5 = layers.Encoder(nout * 6, nout * 8, 3, 2, 1)
+
+ self.aspp = layers.ASPPModule(nout * 8, nout * 8, dilations, dropout=True)
+
+ self.dec4 = layers.Decoder(nout * (6 + 8), nout * 6, 3, 1, 1)
+ self.dec3 = layers.Decoder(nout * (4 + 6), nout * 4, 3, 1, 1)
+ self.dec2 = layers.Decoder(nout * (2 + 4), nout * 2, 3, 1, 1)
+ self.lstm_dec2 = layers.LSTMModule(nout * 2, nin_lstm, nout_lstm)
+ self.dec1 = layers.Decoder(nout * (1 + 2) + 1, nout * 1, 3, 1, 1)
+
+ def __call__(self, x):
+ e1 = self.enc1(x)
+ e2 = self.enc2(e1)
+ e3 = self.enc3(e2)
+ e4 = self.enc4(e3)
+ e5 = self.enc5(e4)
+
+ h = self.aspp(e5)
+
+ h = self.dec4(h, e4)
+ h = self.dec3(h, e3)
+ h = self.dec2(h, e2)
+ h = torch.cat([h, self.lstm_dec2(h)], dim=1)
+ h = self.dec1(h, e1)
+
+ return h
+
+
+class CascadedNet(nn.Module):
+ def __init__(self, n_fft, nout=32, nout_lstm=128):
+ super(CascadedNet, self).__init__()
+
+ self.max_bin = n_fft // 2
+ self.output_bin = n_fft // 2 + 1
+ self.nin_lstm = self.max_bin // 2
+ self.offset = 64
+
+ self.stg1_low_band_net = nn.Sequential(
+ BaseNet(2, nout // 2, self.nin_lstm // 2, nout_lstm),
+ layers.Conv2DBNActiv(nout // 2, nout // 4, 1, 1, 0),
+ )
+
+ self.stg1_high_band_net = BaseNet(
+ 2, nout // 4, self.nin_lstm // 2, nout_lstm // 2
+ )
+
+ self.stg2_low_band_net = nn.Sequential(
+ BaseNet(nout // 4 + 2, nout, self.nin_lstm // 2, nout_lstm),
+ layers.Conv2DBNActiv(nout, nout // 2, 1, 1, 0),
+ )
+ self.stg2_high_band_net = BaseNet(
+ nout // 4 + 2, nout // 2, self.nin_lstm // 2, nout_lstm // 2
+ )
+
+ self.stg3_full_band_net = BaseNet(
+ 3 * nout // 4 + 2, nout, self.nin_lstm, nout_lstm
+ )
+
+ self.out = nn.Conv2d(nout, 2, 1, bias=False)
+ self.aux_out = nn.Conv2d(3 * nout // 4, 2, 1, bias=False)
+
+ def forward(self, x):
+ x = x[:, :, : self.max_bin]
+
+ bandw = x.size()[2] // 2
+ l1_in = x[:, :, :bandw]
+ h1_in = x[:, :, bandw:]
+ l1 = self.stg1_low_band_net(l1_in)
+ h1 = self.stg1_high_band_net(h1_in)
+ aux1 = torch.cat([l1, h1], dim=2)
+
+ l2_in = torch.cat([l1_in, l1], dim=1)
+ h2_in = torch.cat([h1_in, h1], dim=1)
+ l2 = self.stg2_low_band_net(l2_in)
+ h2 = self.stg2_high_band_net(h2_in)
+ aux2 = torch.cat([l2, h2], dim=2)
+
+ f3_in = torch.cat([x, aux1, aux2], dim=1)
+ f3 = self.stg3_full_band_net(f3_in)
+
+ mask = torch.sigmoid(self.out(f3))
+ mask = F.pad(
+ input=mask,
+ pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+ mode="replicate",
+ )
+
+ if self.training:
+ aux = torch.cat([aux1, aux2], dim=1)
+ aux = torch.sigmoid(self.aux_out(aux))
+ aux = F.pad(
+ input=aux,
+ pad=(0, 0, 0, self.output_bin - aux.size()[2]),
+ mode="replicate",
+ )
+ return mask, aux
+ else:
+ return mask
+
+ def predict_mask(self, x):
+ mask = self.forward(x)
+
+ if self.offset > 0:
+ mask = mask[:, :, :, self.offset : -self.offset]
+ assert mask.size()[3] > 0
+
+ return mask
+
+ def predict(self, x, aggressiveness=None):
+ mask = self.forward(x)
+ pred_mag = x * mask
+
+ if self.offset > 0:
+ pred_mag = pred_mag[:, :, :, self.offset : -self.offset]
+ assert pred_mag.size()[3] > 0
+
+ return pred_mag
diff --git a/infer/lib/uvr5_pack/lib_v5/nets_123821KB.py b/infer/lib/uvr5_pack/lib_v5/nets_123821KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..167d4cb2198863cf43e93440f7e63c5342fc7605
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/nets_123821KB.py
@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_123821KB as layers
+
+
+class BaseASPPNet(nn.Module):
+ def __init__(self, nin, ch, dilations=(4, 8, 16)):
+ super(BaseASPPNet, self).__init__()
+ self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+ self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+ self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+ self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+ self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+ self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+ self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+ self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+ self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+ def __call__(self, x):
+ h, e1 = self.enc1(x)
+ h, e2 = self.enc2(h)
+ h, e3 = self.enc3(h)
+ h, e4 = self.enc4(h)
+
+ h = self.aspp(h)
+
+ h = self.dec4(h, e4)
+ h = self.dec3(h, e3)
+ h = self.dec2(h, e2)
+ h = self.dec1(h, e1)
+
+ return h
+
+
+class CascadedASPPNet(nn.Module):
+ def __init__(self, n_fft):
+ super(CascadedASPPNet, self).__init__()
+ self.stg1_low_band_net = BaseASPPNet(2, 32)
+ self.stg1_high_band_net = BaseASPPNet(2, 32)
+
+ self.stg2_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+ self.stg2_full_band_net = BaseASPPNet(16, 32)
+
+ self.stg3_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+ self.stg3_full_band_net = BaseASPPNet(32, 64)
+
+ self.out = nn.Conv2d(64, 2, 1, bias=False)
+ self.aux1_out = nn.Conv2d(32, 2, 1, bias=False)
+ self.aux2_out = nn.Conv2d(32, 2, 1, bias=False)
+
+ self.max_bin = n_fft // 2
+ self.output_bin = n_fft // 2 + 1
+
+ self.offset = 128
+
+ def forward(self, x, aggressiveness=None):
+ mix = x.detach()
+ x = x.clone()
+
+ x = x[:, :, : self.max_bin]
+
+ bandw = x.size()[2] // 2
+ aux1 = torch.cat(
+ [
+ self.stg1_low_band_net(x[:, :, :bandw]),
+ self.stg1_high_band_net(x[:, :, bandw:]),
+ ],
+ dim=2,
+ )
+
+ h = torch.cat([x, aux1], dim=1)
+ aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+ h = torch.cat([x, aux1, aux2], dim=1)
+ h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+ mask = torch.sigmoid(self.out(h))
+ mask = F.pad(
+ input=mask,
+ pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+ mode="replicate",
+ )
+
+ if self.training:
+ aux1 = torch.sigmoid(self.aux1_out(aux1))
+ aux1 = F.pad(
+ input=aux1,
+ pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+ mode="replicate",
+ )
+ aux2 = torch.sigmoid(self.aux2_out(aux2))
+ aux2 = F.pad(
+ input=aux2,
+ pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+ mode="replicate",
+ )
+ return mask * mix, aux1 * mix, aux2 * mix
+ else:
+ if aggressiveness:
+ mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+ mask[:, :, : aggressiveness["split_bin"]],
+ 1 + aggressiveness["value"] / 3,
+ )
+ mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+ mask[:, :, aggressiveness["split_bin"] :],
+ 1 + aggressiveness["value"],
+ )
+
+ return mask * mix
+
+ def predict(self, x_mag, aggressiveness=None):
+ h = self.forward(x_mag, aggressiveness)
+
+ if self.offset > 0:
+ h = h[:, :, :, self.offset : -self.offset]
+ assert h.size()[3] > 0
+
+ return h
diff --git a/infer/lib/uvr5_pack/lib_v5/nets_33966KB.py b/infer/lib/uvr5_pack/lib_v5/nets_33966KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..73a5b836177b706c306e27875f8391c1aed4b948
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/nets_33966KB.py
@@ -0,0 +1,122 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_33966KB as layers
+
+
+class BaseASPPNet(nn.Module):
+ def __init__(self, nin, ch, dilations=(4, 8, 16, 32)):
+ super(BaseASPPNet, self).__init__()
+ self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+ self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+ self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+ self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+ self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+ self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+ self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+ self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+ self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+ def __call__(self, x):
+ h, e1 = self.enc1(x)
+ h, e2 = self.enc2(h)
+ h, e3 = self.enc3(h)
+ h, e4 = self.enc4(h)
+
+ h = self.aspp(h)
+
+ h = self.dec4(h, e4)
+ h = self.dec3(h, e3)
+ h = self.dec2(h, e2)
+ h = self.dec1(h, e1)
+
+ return h
+
+
+class CascadedASPPNet(nn.Module):
+ def __init__(self, n_fft):
+ super(CascadedASPPNet, self).__init__()
+ self.stg1_low_band_net = BaseASPPNet(2, 16)
+ self.stg1_high_band_net = BaseASPPNet(2, 16)
+
+ self.stg2_bridge = layers.Conv2DBNActiv(18, 8, 1, 1, 0)
+ self.stg2_full_band_net = BaseASPPNet(8, 16)
+
+ self.stg3_bridge = layers.Conv2DBNActiv(34, 16, 1, 1, 0)
+ self.stg3_full_band_net = BaseASPPNet(16, 32)
+
+ self.out = nn.Conv2d(32, 2, 1, bias=False)
+ self.aux1_out = nn.Conv2d(16, 2, 1, bias=False)
+ self.aux2_out = nn.Conv2d(16, 2, 1, bias=False)
+
+ self.max_bin = n_fft // 2
+ self.output_bin = n_fft // 2 + 1
+
+ self.offset = 128
+
+ def forward(self, x, aggressiveness=None):
+ mix = x.detach()
+ x = x.clone()
+
+ x = x[:, :, : self.max_bin]
+
+ bandw = x.size()[2] // 2
+ aux1 = torch.cat(
+ [
+ self.stg1_low_band_net(x[:, :, :bandw]),
+ self.stg1_high_band_net(x[:, :, bandw:]),
+ ],
+ dim=2,
+ )
+
+ h = torch.cat([x, aux1], dim=1)
+ aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+ h = torch.cat([x, aux1, aux2], dim=1)
+ h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+ mask = torch.sigmoid(self.out(h))
+ mask = F.pad(
+ input=mask,
+ pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+ mode="replicate",
+ )
+
+ if self.training:
+ aux1 = torch.sigmoid(self.aux1_out(aux1))
+ aux1 = F.pad(
+ input=aux1,
+ pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+ mode="replicate",
+ )
+ aux2 = torch.sigmoid(self.aux2_out(aux2))
+ aux2 = F.pad(
+ input=aux2,
+ pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+ mode="replicate",
+ )
+ return mask * mix, aux1 * mix, aux2 * mix
+ else:
+ if aggressiveness:
+ mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+ mask[:, :, : aggressiveness["split_bin"]],
+ 1 + aggressiveness["value"] / 3,
+ )
+ mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+ mask[:, :, aggressiveness["split_bin"] :],
+ 1 + aggressiveness["value"],
+ )
+
+ return mask * mix
+
+ def predict(self, x_mag, aggressiveness=None):
+ h = self.forward(x_mag, aggressiveness)
+
+ if self.offset > 0:
+ h = h[:, :, :, self.offset : -self.offset]
+ assert h.size()[3] > 0
+
+ return h
diff --git a/infer/lib/uvr5_pack/lib_v5/nets_537238KB.py b/infer/lib/uvr5_pack/lib_v5/nets_537238KB.py
new file mode 100644
index 0000000000000000000000000000000000000000..823b44fb64898e8dcbb12180ba45d1718f9b03f7
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/nets_537238KB.py
@@ -0,0 +1,123 @@
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from . import layers_537238KB as layers
+
+
+class BaseASPPNet(nn.Module):
+ def __init__(self, nin, ch, dilations=(4, 8, 16)):
+ super(BaseASPPNet, self).__init__()
+ self.enc1 = layers.Encoder(nin, ch, 3, 2, 1)
+ self.enc2 = layers.Encoder(ch, ch * 2, 3, 2, 1)
+ self.enc3 = layers.Encoder(ch * 2, ch * 4, 3, 2, 1)
+ self.enc4 = layers.Encoder(ch * 4, ch * 8, 3, 2, 1)
+
+ self.aspp = layers.ASPPModule(ch * 8, ch * 16, dilations)
+
+ self.dec4 = layers.Decoder(ch * (8 + 16), ch * 8, 3, 1, 1)
+ self.dec3 = layers.Decoder(ch * (4 + 8), ch * 4, 3, 1, 1)
+ self.dec2 = layers.Decoder(ch * (2 + 4), ch * 2, 3, 1, 1)
+ self.dec1 = layers.Decoder(ch * (1 + 2), ch, 3, 1, 1)
+
+ def __call__(self, x):
+ h, e1 = self.enc1(x)
+ h, e2 = self.enc2(h)
+ h, e3 = self.enc3(h)
+ h, e4 = self.enc4(h)
+
+ h = self.aspp(h)
+
+ h = self.dec4(h, e4)
+ h = self.dec3(h, e3)
+ h = self.dec2(h, e2)
+ h = self.dec1(h, e1)
+
+ return h
+
+
+class CascadedASPPNet(nn.Module):
+ def __init__(self, n_fft):
+ super(CascadedASPPNet, self).__init__()
+ self.stg1_low_band_net = BaseASPPNet(2, 64)
+ self.stg1_high_band_net = BaseASPPNet(2, 64)
+
+ self.stg2_bridge = layers.Conv2DBNActiv(66, 32, 1, 1, 0)
+ self.stg2_full_band_net = BaseASPPNet(32, 64)
+
+ self.stg3_bridge = layers.Conv2DBNActiv(130, 64, 1, 1, 0)
+ self.stg3_full_band_net = BaseASPPNet(64, 128)
+
+ self.out = nn.Conv2d(128, 2, 1, bias=False)
+ self.aux1_out = nn.Conv2d(64, 2, 1, bias=False)
+ self.aux2_out = nn.Conv2d(64, 2, 1, bias=False)
+
+ self.max_bin = n_fft // 2
+ self.output_bin = n_fft // 2 + 1
+
+ self.offset = 128
+
+ def forward(self, x, aggressiveness=None):
+ mix = x.detach()
+ x = x.clone()
+
+ x = x[:, :, : self.max_bin]
+
+ bandw = x.size()[2] // 2
+ aux1 = torch.cat(
+ [
+ self.stg1_low_band_net(x[:, :, :bandw]),
+ self.stg1_high_band_net(x[:, :, bandw:]),
+ ],
+ dim=2,
+ )
+
+ h = torch.cat([x, aux1], dim=1)
+ aux2 = self.stg2_full_band_net(self.stg2_bridge(h))
+
+ h = torch.cat([x, aux1, aux2], dim=1)
+ h = self.stg3_full_band_net(self.stg3_bridge(h))
+
+ mask = torch.sigmoid(self.out(h))
+ mask = F.pad(
+ input=mask,
+ pad=(0, 0, 0, self.output_bin - mask.size()[2]),
+ mode="replicate",
+ )
+
+ if self.training:
+ aux1 = torch.sigmoid(self.aux1_out(aux1))
+ aux1 = F.pad(
+ input=aux1,
+ pad=(0, 0, 0, self.output_bin - aux1.size()[2]),
+ mode="replicate",
+ )
+ aux2 = torch.sigmoid(self.aux2_out(aux2))
+ aux2 = F.pad(
+ input=aux2,
+ pad=(0, 0, 0, self.output_bin - aux2.size()[2]),
+ mode="replicate",
+ )
+ return mask * mix, aux1 * mix, aux2 * mix
+ else:
+ if aggressiveness:
+ mask[:, :, : aggressiveness["split_bin"]] = torch.pow(
+ mask[:, :, : aggressiveness["split_bin"]],
+ 1 + aggressiveness["value"] / 3,
+ )
+ mask[:, :, aggressiveness["split_bin"] :] = torch.pow(
+ mask[:, :, aggressiveness["split_bin"] :],
+ 1 + aggressiveness["value"],
+ )
+
+ return mask * mix
+
+ def predict(self, x_mag, aggressiveness=None):
+ h = self.forward(x_mag, aggressiveness)
+
+ if self.offset > 0:
+ h = h[:, :, :, self.offset : -self.offset]
+ assert h.size()[3] > 0
+
+ return h
diff --git a/infer/lib/uvr5_pack/lib_v5/spec_utils.py b/infer/lib/uvr5_pack/lib_v5/spec_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4637635ee1d69fe4a2df393f00f7b5066112fad9
--- /dev/null
+++ b/infer/lib/uvr5_pack/lib_v5/spec_utils.py
@@ -0,0 +1,520 @@
+import hashlib
+import json
+import math
+import os
+
+import librosa
+import numpy as np
+
+
+def crop_center(h1, h2):
+ h1_shape = h1.size()
+ h2_shape = h2.size()
+
+ if h1_shape[3] == h2_shape[3]:
+ return h1
+ elif h1_shape[3] < h2_shape[3]:
+ raise ValueError("h1_shape[3] must be greater than h2_shape[3]")
+
+ # s_freq = (h2_shape[2] - h1_shape[2]) // 2
+ # e_freq = s_freq + h1_shape[2]
+ s_time = (h1_shape[3] - h2_shape[3]) // 2
+ e_time = s_time + h2_shape[3]
+ h1 = h1[:, :, :, s_time:e_time]
+
+ return h1
+
+
+def wave_to_spectrogram(
+ wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
+):
+ if reverse:
+ wave_left = np.flip(np.asfortranarray(wave[0]))
+ wave_right = np.flip(np.asfortranarray(wave[1]))
+ elif mid_side:
+ wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+ wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+ elif mid_side_b2:
+ wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
+ wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
+ else:
+ wave_left = np.asfortranarray(wave[0])
+ wave_right = np.asfortranarray(wave[1])
+
+ spec_left = librosa.stft(wave_left, n_fft, hop_length=hop_length)
+ spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+
+ spec = np.asfortranarray([spec_left, spec_right])
+
+ return spec
+
+
+def wave_to_spectrogram_mt(
+ wave, hop_length, n_fft, mid_side=False, mid_side_b2=False, reverse=False
+):
+ import threading
+
+ if reverse:
+ wave_left = np.flip(np.asfortranarray(wave[0]))
+ wave_right = np.flip(np.asfortranarray(wave[1]))
+ elif mid_side:
+ wave_left = np.asfortranarray(np.add(wave[0], wave[1]) / 2)
+ wave_right = np.asfortranarray(np.subtract(wave[0], wave[1]))
+ elif mid_side_b2:
+ wave_left = np.asfortranarray(np.add(wave[1], wave[0] * 0.5))
+ wave_right = np.asfortranarray(np.subtract(wave[0], wave[1] * 0.5))
+ else:
+ wave_left = np.asfortranarray(wave[0])
+ wave_right = np.asfortranarray(wave[1])
+
+ def run_thread(**kwargs):
+ global spec_left
+ spec_left = librosa.stft(**kwargs)
+
+ thread = threading.Thread(
+ target=run_thread,
+ kwargs={"y": wave_left, "n_fft": n_fft, "hop_length": hop_length},
+ )
+ thread.start()
+ spec_right = librosa.stft(wave_right, n_fft, hop_length=hop_length)
+ thread.join()
+
+ spec = np.asfortranarray([spec_left, spec_right])
+
+ return spec
+
+
+def combine_spectrograms(specs, mp):
+ l = min([specs[i].shape[2] for i in specs])
+ spec_c = np.zeros(shape=(2, mp.param["bins"] + 1, l), dtype=np.complex64)
+ offset = 0
+ bands_n = len(mp.param["band"])
+
+ for d in range(1, bands_n + 1):
+ h = mp.param["band"][d]["crop_stop"] - mp.param["band"][d]["crop_start"]
+ spec_c[:, offset : offset + h, :l] = specs[d][
+ :, mp.param["band"][d]["crop_start"] : mp.param["band"][d]["crop_stop"], :l
+ ]
+ offset += h
+
+ if offset > mp.param["bins"]:
+ raise ValueError("Too much bins")
+
+ # lowpass fiter
+ if (
+ mp.param["pre_filter_start"] > 0
+ ): # and mp.param['band'][bands_n]['res_type'] in ['scipy', 'polyphase']:
+ if bands_n == 1:
+ spec_c = fft_lp_filter(
+ spec_c, mp.param["pre_filter_start"], mp.param["pre_filter_stop"]
+ )
+ else:
+ gp = 1
+ for b in range(
+ mp.param["pre_filter_start"] + 1, mp.param["pre_filter_stop"]
+ ):
+ g = math.pow(
+ 10, -(b - mp.param["pre_filter_start"]) * (3.5 - gp) / 20.0
+ )
+ gp = g
+ spec_c[:, b, :] *= g
+
+ return np.asfortranarray(spec_c)
+
+
+def spectrogram_to_image(spec, mode="magnitude"):
+ if mode == "magnitude":
+ if np.iscomplexobj(spec):
+ y = np.abs(spec)
+ else:
+ y = spec
+ y = np.log10(y**2 + 1e-8)
+ elif mode == "phase":
+ if np.iscomplexobj(spec):
+ y = np.angle(spec)
+ else:
+ y = spec
+
+ y -= y.min()
+ y *= 255 / y.max()
+ img = np.uint8(y)
+
+ if y.ndim == 3:
+ img = img.transpose(1, 2, 0)
+ img = np.concatenate([np.max(img, axis=2, keepdims=True), img], axis=2)
+
+ return img
+
+
+def reduce_vocal_aggressively(X, y, softmask):
+ v = X - y
+ y_mag_tmp = np.abs(y)
+ v_mag_tmp = np.abs(v)
+
+ v_mask = v_mag_tmp > y_mag_tmp
+ y_mag = np.clip(y_mag_tmp - v_mag_tmp * v_mask * softmask, 0, np.inf)
+
+ return y_mag * np.exp(1.0j * np.angle(y))
+
+
+def mask_silence(mag, ref, thres=0.2, min_range=64, fade_size=32):
+ if min_range < fade_size * 2:
+ raise ValueError("min_range must be >= fade_area * 2")
+
+ mag = mag.copy()
+
+ idx = np.where(ref.mean(axis=(0, 1)) < thres)[0]
+ starts = np.insert(idx[np.where(np.diff(idx) != 1)[0] + 1], 0, idx[0])
+ ends = np.append(idx[np.where(np.diff(idx) != 1)[0]], idx[-1])
+ uninformative = np.where(ends - starts > min_range)[0]
+ if len(uninformative) > 0:
+ starts = starts[uninformative]
+ ends = ends[uninformative]
+ old_e = None
+ for s, e in zip(starts, ends):
+ if old_e is not None and s - old_e < fade_size:
+ s = old_e - fade_size * 2
+
+ if s != 0:
+ weight = np.linspace(0, 1, fade_size)
+ mag[:, :, s : s + fade_size] += weight * ref[:, :, s : s + fade_size]
+ else:
+ s -= fade_size
+
+ if e != mag.shape[2]:
+ weight = np.linspace(1, 0, fade_size)
+ mag[:, :, e - fade_size : e] += weight * ref[:, :, e - fade_size : e]
+ else:
+ e += fade_size
+
+ mag[:, :, s + fade_size : e - fade_size] += ref[
+ :, :, s + fade_size : e - fade_size
+ ]
+ old_e = e
+
+ return mag
+
+
+def align_wave_head_and_tail(a, b):
+ l = min([a[0].size, b[0].size])
+
+ return a[:l, :l], b[:l, :l]
+
+
+def cache_or_load(mix_path, inst_path, mp):
+ mix_basename = os.path.splitext(os.path.basename(mix_path))[0]
+ inst_basename = os.path.splitext(os.path.basename(inst_path))[0]
+
+ cache_dir = "mph{}".format(
+ hashlib.sha1(json.dumps(mp.param, sort_keys=True).encode("utf-8")).hexdigest()
+ )
+ mix_cache_dir = os.path.join("cache", cache_dir)
+ inst_cache_dir = os.path.join("cache", cache_dir)
+
+ os.makedirs(mix_cache_dir, exist_ok=True)
+ os.makedirs(inst_cache_dir, exist_ok=True)
+
+ mix_cache_path = os.path.join(mix_cache_dir, mix_basename + ".npy")
+ inst_cache_path = os.path.join(inst_cache_dir, inst_basename + ".npy")
+
+ if os.path.exists(mix_cache_path) and os.path.exists(inst_cache_path):
+ X_spec_m = np.load(mix_cache_path)
+ y_spec_m = np.load(inst_cache_path)
+ else:
+ X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+
+ for d in range(len(mp.param["band"]), 0, -1):
+ bp = mp.param["band"][d]
+
+ if d == len(mp.param["band"]): # high-end band
+ X_wave[d], _ = librosa.load(
+ mix_path, bp["sr"], False, dtype=np.float32, res_type=bp["res_type"]
+ )
+ y_wave[d], _ = librosa.load(
+ inst_path,
+ bp["sr"],
+ False,
+ dtype=np.float32,
+ res_type=bp["res_type"],
+ )
+ else: # lower bands
+ X_wave[d] = librosa.resample(
+ X_wave[d + 1],
+ mp.param["band"][d + 1]["sr"],
+ bp["sr"],
+ res_type=bp["res_type"],
+ )
+ y_wave[d] = librosa.resample(
+ y_wave[d + 1],
+ mp.param["band"][d + 1]["sr"],
+ bp["sr"],
+ res_type=bp["res_type"],
+ )
+
+ X_wave[d], y_wave[d] = align_wave_head_and_tail(X_wave[d], y_wave[d])
+
+ X_spec_s[d] = wave_to_spectrogram(
+ X_wave[d],
+ bp["hl"],
+ bp["n_fft"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ )
+ y_spec_s[d] = wave_to_spectrogram(
+ y_wave[d],
+ bp["hl"],
+ bp["n_fft"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ )
+
+ del X_wave, y_wave
+
+ X_spec_m = combine_spectrograms(X_spec_s, mp)
+ y_spec_m = combine_spectrograms(y_spec_s, mp)
+
+ if X_spec_m.shape != y_spec_m.shape:
+ raise ValueError("The combined spectrograms are different: " + mix_path)
+
+ _, ext = os.path.splitext(mix_path)
+
+ np.save(mix_cache_path, X_spec_m)
+ np.save(inst_cache_path, y_spec_m)
+
+ return X_spec_m, y_spec_m
+
+
+def spectrogram_to_wave(spec, hop_length, mid_side, mid_side_b2, reverse):
+ spec_left = np.asfortranarray(spec[0])
+ spec_right = np.asfortranarray(spec[1])
+
+ wave_left = librosa.istft(spec_left, hop_length=hop_length)
+ wave_right = librosa.istft(spec_right, hop_length=hop_length)
+
+ if reverse:
+ return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+ elif mid_side:
+ return np.asfortranarray(
+ [np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
+ )
+ elif mid_side_b2:
+ return np.asfortranarray(
+ [
+ np.add(wave_right / 1.25, 0.4 * wave_left),
+ np.subtract(wave_left / 1.25, 0.4 * wave_right),
+ ]
+ )
+ else:
+ return np.asfortranarray([wave_left, wave_right])
+
+
+def spectrogram_to_wave_mt(spec, hop_length, mid_side, reverse, mid_side_b2):
+ import threading
+
+ spec_left = np.asfortranarray(spec[0])
+ spec_right = np.asfortranarray(spec[1])
+
+ def run_thread(**kwargs):
+ global wave_left
+ wave_left = librosa.istft(**kwargs)
+
+ thread = threading.Thread(
+ target=run_thread, kwargs={"stft_matrix": spec_left, "hop_length": hop_length}
+ )
+ thread.start()
+ wave_right = librosa.istft(spec_right, hop_length=hop_length)
+ thread.join()
+
+ if reverse:
+ return np.asfortranarray([np.flip(wave_left), np.flip(wave_right)])
+ elif mid_side:
+ return np.asfortranarray(
+ [np.add(wave_left, wave_right / 2), np.subtract(wave_left, wave_right / 2)]
+ )
+ elif mid_side_b2:
+ return np.asfortranarray(
+ [
+ np.add(wave_right / 1.25, 0.4 * wave_left),
+ np.subtract(wave_left / 1.25, 0.4 * wave_right),
+ ]
+ )
+ else:
+ return np.asfortranarray([wave_left, wave_right])
+
+
+def cmb_spectrogram_to_wave(spec_m, mp, extra_bins_h=None, extra_bins=None):
+ wave_band = {}
+ bands_n = len(mp.param["band"])
+ offset = 0
+
+ for d in range(1, bands_n + 1):
+ bp = mp.param["band"][d]
+ spec_s = np.ndarray(
+ shape=(2, bp["n_fft"] // 2 + 1, spec_m.shape[2]), dtype=complex
+ )
+ h = bp["crop_stop"] - bp["crop_start"]
+ spec_s[:, bp["crop_start"] : bp["crop_stop"], :] = spec_m[
+ :, offset : offset + h, :
+ ]
+
+ offset += h
+ if d == bands_n: # higher
+ if extra_bins_h: # if --high_end_process bypass
+ max_bin = bp["n_fft"] // 2
+ spec_s[:, max_bin - extra_bins_h : max_bin, :] = extra_bins[
+ :, :extra_bins_h, :
+ ]
+ if bp["hpf_start"] > 0:
+ spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
+ if bands_n == 1:
+ wave = spectrogram_to_wave(
+ spec_s,
+ bp["hl"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ )
+ else:
+ wave = np.add(
+ wave,
+ spectrogram_to_wave(
+ spec_s,
+ bp["hl"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ ),
+ )
+ else:
+ sr = mp.param["band"][d + 1]["sr"]
+ if d == 1: # lower
+ spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
+ wave = librosa.resample(
+ spectrogram_to_wave(
+ spec_s,
+ bp["hl"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ ),
+ bp["sr"],
+ sr,
+ res_type="sinc_fastest",
+ )
+ else: # mid
+ spec_s = fft_hp_filter(spec_s, bp["hpf_start"], bp["hpf_stop"] - 1)
+ spec_s = fft_lp_filter(spec_s, bp["lpf_start"], bp["lpf_stop"])
+ wave2 = np.add(
+ wave,
+ spectrogram_to_wave(
+ spec_s,
+ bp["hl"],
+ mp.param["mid_side"],
+ mp.param["mid_side_b2"],
+ mp.param["reverse"],
+ ),
+ )
+ # wave = librosa.core.resample(wave2, bp['sr'], sr, res_type="sinc_fastest")
+ wave = librosa.core.resample(wave2, bp["sr"], sr, res_type="scipy")
+
+ return wave.T
+
+
+def fft_lp_filter(spec, bin_start, bin_stop):
+ g = 1.0
+ for b in range(bin_start, bin_stop):
+ g -= 1 / (bin_stop - bin_start)
+ spec[:, b, :] = g * spec[:, b, :]
+
+ spec[:, bin_stop:, :] *= 0
+
+ return spec
+
+
+def fft_hp_filter(spec, bin_start, bin_stop):
+ g = 1.0
+ for b in range(bin_start, bin_stop, -1):
+ g -= 1 / (bin_start - bin_stop)
+ spec[:, b, :] = g * spec[:, b, :]
+
+ spec[:, 0 : bin_stop + 1, :] *= 0
+
+ return spec
+
+
+def mirroring(a, spec_m, input_high_end, mp):
+ if "mirroring" == a:
+ mirror = np.flip(
+ np.abs(
+ spec_m[
+ :,
+ mp.param["pre_filter_start"]
+ - 10
+ - input_high_end.shape[1] : mp.param["pre_filter_start"]
+ - 10,
+ :,
+ ]
+ ),
+ 1,
+ )
+ mirror = mirror * np.exp(1.0j * np.angle(input_high_end))
+
+ return np.where(
+ np.abs(input_high_end) <= np.abs(mirror), input_high_end, mirror
+ )
+
+ if "mirroring2" == a:
+ mirror = np.flip(
+ np.abs(
+ spec_m[
+ :,
+ mp.param["pre_filter_start"]
+ - 10
+ - input_high_end.shape[1] : mp.param["pre_filter_start"]
+ - 10,
+ :,
+ ]
+ ),
+ 1,
+ )
+ mi = np.multiply(mirror, input_high_end * 1.7)
+
+ return np.where(np.abs(input_high_end) <= np.abs(mi), input_high_end, mi)
+
+
+def ensembling(a, specs):
+ for i in range(1, len(specs)):
+ if i == 1:
+ spec = specs[0]
+
+ ln = min([spec.shape[2], specs[i].shape[2]])
+ spec = spec[:, :, :ln]
+ specs[i] = specs[i][:, :, :ln]
+
+ if "min_mag" == a:
+ spec = np.where(np.abs(specs[i]) <= np.abs(spec), specs[i], spec)
+ if "max_mag" == a:
+ spec = np.where(np.abs(specs[i]) >= np.abs(spec), specs[i], spec)
+
+ return spec
+
+
+def stft(wave, nfft, hl):
+ wave_left = np.asfortranarray(wave[0])
+ wave_right = np.asfortranarray(wave[1])
+ spec_left = librosa.stft(wave_left, nfft, hop_length=hl)
+ spec_right = librosa.stft(wave_right, nfft, hop_length=hl)
+ spec = np.asfortranarray([spec_left, spec_right])
+
+ return spec
+
+
+def istft(spec, hl):
+ spec_left = np.asfortranarray(spec[0])
+ spec_right = np.asfortranarray(spec[1])
+
+ wave_left = librosa.istft(spec_left, hop_length=hl)
+ wave_right = librosa.istft(spec_right, hop_length=hl)
+ wave = np.asfortranarray([wave_left, wave_right])
diff --git a/infer/lib/uvr5_pack/name_params.json b/infer/lib/uvr5_pack/name_params.json
new file mode 100644
index 0000000000000000000000000000000000000000..8ed51a68370607a7a8693b99cfb35fc5d92b04af
--- /dev/null
+++ b/infer/lib/uvr5_pack/name_params.json
@@ -0,0 +1,263 @@
+{
+ "equivalent" : [
+ {
+ "model_hash_name" : [
+ {
+ "hash_name": "47939caf0cfe52a0e81442b85b971dfd",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "4e4ecb9764c50a8c414fee6e10395bbe",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+ "param_name": "4band_v2"
+ },
+ {
+ "hash_name": "ca106edd563e034bde0bdec4bb7a4b36",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+ "param_name": "4band_v2"
+ },
+ {
+ "hash_name": "e60a1e84803ce4efc0a6551206cc4b71",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "a82f14e75892e55e994376edbf0c8435",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "6dd9eaa6f0420af9f1d403aaafa4cc06",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+ "param_name": "4band_v2_sn"
+ },
+ {
+ "hash_name": "08611fb99bd59eaa79ad27c58d137727",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+ "param_name": "4band_v2_sn"
+ },
+ {
+ "hash_name": "5c7bbca45a187e81abbbd351606164e5",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+ "param_name": "3band_44100_msb2"
+ },
+ {
+ "hash_name": "d6b2cb685a058a091e5e7098192d3233",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+ "param_name": "3band_44100_msb2"
+ },
+ {
+ "hash_name": "c1b9f38170a7c90e96f027992eb7c62b",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "c3448ec923fa0edf3d03a19e633faa53",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "68aa2c8093d0080704b200d140f59e54",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json",
+ "param_name": "3band_44100"
+ },
+ {
+ "hash_name": "fdc83be5b798e4bd29fe00fe6600e147",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+ "param_name": "3band_44100_mid.json"
+ },
+ {
+ "hash_name": "2ce34bc92fd57f55db16b7a4def3d745",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+ "param_name": "3band_44100_mid.json"
+ },
+ {
+ "hash_name": "52fdca89576f06cf4340b74a4730ee5f",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100.json"
+ },
+ {
+ "hash_name": "41191165b05d38fc77f072fa9e8e8a30",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100.json"
+ },
+ {
+ "hash_name": "89e83b511ad474592689e562d5b1f80e",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+ "param_name": "2band_32000.json"
+ },
+ {
+ "hash_name": "0b954da81d453b716b114d6d7c95177f",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+ "param_name": "2band_32000.json"
+ }
+
+ ],
+ "v4 Models": [
+ {
+ "hash_name": "6a00461c51c2920fd68937d4609ed6c8",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
+ "param_name": "1band_sr16000_hl512"
+ },
+ {
+ "hash_name": "0ab504864d20f1bd378fe9c81ef37140",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+ "param_name": "1band_sr32000_hl512"
+ },
+ {
+ "hash_name": "7dd21065bf91c10f7fccb57d7d83b07f",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+ "param_name": "1band_sr32000_hl512"
+ },
+ {
+ "hash_name": "80ab74d65e515caa3622728d2de07d23",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+ "param_name": "1band_sr32000_hl512"
+ },
+ {
+ "hash_name": "edc115e7fc523245062200c00caa847f",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+ "param_name": "1band_sr33075_hl384"
+ },
+ {
+ "hash_name": "28063e9f6ab5b341c5f6d3c67f2045b7",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+ "param_name": "1band_sr33075_hl384"
+ },
+ {
+ "hash_name": "b58090534c52cbc3e9b5104bad666ef2",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+ "param_name": "1band_sr44100_hl512"
+ },
+ {
+ "hash_name": "0cdab9947f1b0928705f518f3c78ea8f",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+ "param_name": "1band_sr44100_hl512"
+ },
+ {
+ "hash_name": "ae702fed0238afb5346db8356fe25f13",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
+ "param_name": "1band_sr44100_hl1024"
+ }
+ ]
+ }
+ ],
+ "User Models" : [
+ {
+ "1 Band": [
+ {
+ "hash_name": "1band_sr16000_hl512",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr16000_hl512.json",
+ "param_name": "1band_sr16000_hl512"
+ },
+ {
+ "hash_name": "1band_sr32000_hl512",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr32000_hl512.json",
+ "param_name": "1band_sr16000_hl512"
+ },
+ {
+ "hash_name": "1band_sr33075_hl384",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr33075_hl384.json",
+ "param_name": "1band_sr33075_hl384"
+ },
+ {
+ "hash_name": "1band_sr44100_hl256",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl256.json",
+ "param_name": "1band_sr44100_hl256"
+ },
+ {
+ "hash_name": "1band_sr44100_hl512",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl512.json",
+ "param_name": "1band_sr44100_hl512"
+ },
+ {
+ "hash_name": "1band_sr44100_hl1024",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/1band_sr44100_hl1024.json",
+ "param_name": "1band_sr44100_hl1024"
+ }
+ ],
+ "2 Band": [
+ {
+ "hash_name": "2band_44100_lofi",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_44100_lofi.json",
+ "param_name": "2band_44100_lofi"
+ },
+ {
+ "hash_name": "2band_32000",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_32000.json",
+ "param_name": "2band_32000"
+ },
+ {
+ "hash_name": "2band_48000",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/2band_48000.json",
+ "param_name": "2band_48000"
+ }
+ ],
+ "3 Band": [
+ {
+ "hash_name": "3band_44100",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100.json",
+ "param_name": "3band_44100"
+ },
+ {
+ "hash_name": "3band_44100_mid",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_mid.json",
+ "param_name": "3band_44100_mid"
+ },
+ {
+ "hash_name": "3band_44100_msb2",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/3band_44100_msb2.json",
+ "param_name": "3band_44100_msb2"
+ }
+ ],
+ "4 Band": [
+ {
+ "hash_name": "4band_44100",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100.json",
+ "param_name": "4band_44100"
+ },
+ {
+ "hash_name": "4band_44100_mid",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_mid.json",
+ "param_name": "4band_44100_mid"
+ },
+ {
+ "hash_name": "4band_44100_msb",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb.json",
+ "param_name": "4band_44100_msb"
+ },
+ {
+ "hash_name": "4band_44100_msb2",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_msb2.json",
+ "param_name": "4band_44100_msb2"
+ },
+ {
+ "hash_name": "4band_44100_reverse",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_reverse.json",
+ "param_name": "4band_44100_reverse"
+ },
+ {
+ "hash_name": "4band_44100_sw",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_44100_sw.json",
+ "param_name": "4band_44100_sw"
+ },
+ {
+ "hash_name": "4band_v2",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json",
+ "param_name": "4band_v2"
+ },
+ {
+ "hash_name": "4band_v2_sn",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2_sn.json",
+ "param_name": "4band_v2_sn"
+ },
+ {
+ "hash_name": "tmodelparam",
+ "model_params": "infer/lib/uvr5_pack/lib_v5/modelparams/tmodelparam.json",
+ "param_name": "User Model Param Set"
+ }
+ ]
+ }
+ ]
+}
\ No newline at end of file
diff --git a/infer/lib/uvr5_pack/utils.py b/infer/lib/uvr5_pack/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..71056c01e443a07bfb3aae651102565525d9c2a3
--- /dev/null
+++ b/infer/lib/uvr5_pack/utils.py
@@ -0,0 +1,90 @@
+import numpy as np
+import torch
+from tqdm import tqdm
+
+
+def make_padding(width, cropsize, offset):
+ left = offset
+ roi_size = cropsize - left * 2
+ if roi_size == 0:
+ roi_size = cropsize
+ right = roi_size - (width % roi_size) + left
+
+ return left, right, roi_size
+
+
+def inference(X_spec, device, model, aggressiveness, data):
+ """
+ data ： dic configs
+ """
+
+ def _execute(
+ X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half=True
+ ):
+ model.eval()
+ with torch.no_grad():
+ preds = []
+
+ iterations = [n_window]
+
+ total_iterations = sum(iterations)
+ for i in tqdm(range(n_window)):
+ start = i * roi_size
+ X_mag_window = X_mag_pad[
+ None, :, :, start : start + data["window_size"]
+ ]
+ X_mag_window = torch.from_numpy(X_mag_window)
+ if is_half:
+ X_mag_window = X_mag_window.half()
+ X_mag_window = X_mag_window.to(device)
+
+ pred = model.predict(X_mag_window, aggressiveness)
+
+ pred = pred.detach().cpu().numpy()
+ preds.append(pred[0])
+
+ pred = np.concatenate(preds, axis=2)
+ return pred
+
+ def preprocess(X_spec):
+ X_mag = np.abs(X_spec)
+ X_phase = np.angle(X_spec)
+
+ return X_mag, X_phase
+
+ X_mag, X_phase = preprocess(X_spec)
+
+ coef = X_mag.max()
+ X_mag_pre = X_mag / coef
+
+ n_frame = X_mag_pre.shape[2]
+ pad_l, pad_r, roi_size = make_padding(n_frame, data["window_size"], model.offset)
+ n_window = int(np.ceil(n_frame / roi_size))
+
+ X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
+
+ if list(model.state_dict().values())[0].dtype == torch.float16:
+ is_half = True
+ else:
+ is_half = False
+ pred = _execute(
+ X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
+ )
+ pred = pred[:, :, :n_frame]
+
+ if data["tta"]:
+ pad_l += roi_size // 2
+ pad_r += roi_size // 2
+ n_window += 1
+
+ X_mag_pad = np.pad(X_mag_pre, ((0, 0), (0, 0), (pad_l, pad_r)), mode="constant")
+
+ pred_tta = _execute(
+ X_mag_pad, roi_size, n_window, device, model, aggressiveness, is_half
+ )
+ pred_tta = pred_tta[:, :, roi_size // 2 :]
+ pred_tta = pred_tta[:, :, :n_frame]
+
+ return (pred + pred_tta) * 0.5 * coef, X_mag, np.exp(1.0j * X_phase)
+ else:
+ return pred * coef, X_mag, np.exp(1.0j * X_phase)
diff --git a/infer/modules/gui/__init__.py b/infer/modules/gui/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8c1b5494ab8d2d3e932b7ac666121a0c5cf9c81b
--- /dev/null
+++ b/infer/modules/gui/__init__.py
@@ -0,0 +1,13 @@
+"""
+TorchGating is a PyTorch-based implementation of Spectral Gating
+================================================
+Author: Asaf Zorea
+
+Contents
+--------
+torchgate imports all the functions from PyTorch, and in addition provides:
+ TorchGating --- A PyTorch module that applies a spectral gate to an input signal
+
+"""
+
+from .torchgate import TorchGate
diff --git a/infer/modules/gui/torchgate.py b/infer/modules/gui/torchgate.py
new file mode 100644
index 0000000000000000000000000000000000000000..3111bfa72916feb1a340d9e419adb45cf72563a5
--- /dev/null
+++ b/infer/modules/gui/torchgate.py
@@ -0,0 +1,280 @@
+import torch
+from rvc.f0.stft import STFT
+from torch.nn.functional import conv1d, conv2d
+from typing import Union, Optional
+from .utils import linspace, temperature_sigmoid, amp_to_db
+
+
+class TorchGate(torch.nn.Module):
+ """
+ A PyTorch module that applies a spectral gate to an input signal.
+
+ Arguments:
+ sr {int} -- Sample rate of the input signal.
+ nonstationary {bool} -- Whether to use non-stationary or stationary masking (default: {False}).
+ n_std_thresh_stationary {float} -- Number of standard deviations above mean to threshold noise for
+ stationary masking (default: {1.5}).
+ n_thresh_nonstationary {float} -- Number of multiplies above smoothed magnitude spectrogram. for
+ non-stationary masking (default: {1.3}).
+ temp_coeff_nonstationary {float} -- Temperature coefficient for non-stationary masking (default: {0.1}).
+ n_movemean_nonstationary {int} -- Number of samples for moving average smoothing in non-stationary masking
+ (default: {20}).
+ prop_decrease {float} -- Proportion to decrease signal by where the mask is zero (default: {1.0}).
+ n_fft {int} -- Size of FFT for STFT (default: {1024}).
+ win_length {[int]} -- Window length for STFT. If None, defaults to `n_fft` (default: {None}).
+ hop_length {[int]} -- Hop length for STFT. If None, defaults to `win_length` // 4 (default: {None}).
+ freq_mask_smooth_hz {float} -- Frequency smoothing width for mask (in Hz). If None, no smoothing is applied
+ (default: {500}).
+ time_mask_smooth_ms {float} -- Time smoothing width for mask (in ms). If None, no smoothing is applied
+ (default: {50}).
+ """
+
+ @torch.no_grad()
+ def __init__(
+ self,
+ sr: int,
+ nonstationary: bool = False,
+ n_std_thresh_stationary: float = 1.5,
+ n_thresh_nonstationary: float = 1.3,
+ temp_coeff_nonstationary: float = 0.1,
+ n_movemean_nonstationary: int = 20,
+ prop_decrease: float = 1.0,
+ n_fft: int = 1024,
+ win_length: bool = None,
+ hop_length: int = None,
+ freq_mask_smooth_hz: float = 500,
+ time_mask_smooth_ms: float = 50,
+ ):
+ super().__init__()
+
+ # General Params
+ self.sr = sr
+ self.nonstationary = nonstationary
+ assert 0.0 <= prop_decrease <= 1.0
+ self.prop_decrease = prop_decrease
+
+ # STFT Params
+ self.n_fft = n_fft
+ self.win_length = self.n_fft if win_length is None else win_length
+ self.hop_length = self.win_length // 4 if hop_length is None else hop_length
+
+ # Stationary Params
+ self.n_std_thresh_stationary = n_std_thresh_stationary
+
+ # Non-Stationary Params
+ self.temp_coeff_nonstationary = temp_coeff_nonstationary
+ self.n_movemean_nonstationary = n_movemean_nonstationary
+ self.n_thresh_nonstationary = n_thresh_nonstationary
+
+ # Smooth Mask Params
+ self.freq_mask_smooth_hz = freq_mask_smooth_hz
+ self.time_mask_smooth_ms = time_mask_smooth_ms
+ self.register_buffer("smoothing_filter", self._generate_mask_smoothing_filter())
+
+ @torch.no_grad()
+ def _generate_mask_smoothing_filter(self) -> Union[torch.Tensor, None]:
+ """
+ A PyTorch module that applies a spectral gate to an input signal using the STFT.
+
+ Returns:
+ smoothing_filter (torch.Tensor): a 2D tensor representing the smoothing filter,
+ with shape (n_grad_freq, n_grad_time), where n_grad_freq is the number of frequency
+ bins to smooth and n_grad_time is the number of time frames to smooth.
+ If both self.freq_mask_smooth_hz and self.time_mask_smooth_ms are None, returns None.
+ """
+ if self.freq_mask_smooth_hz is None and self.time_mask_smooth_ms is None:
+ return None
+
+ n_grad_freq = (
+ 1
+ if self.freq_mask_smooth_hz is None
+ else int(self.freq_mask_smooth_hz / (self.sr / (self.n_fft / 2)))
+ )
+ if n_grad_freq < 1:
+ raise ValueError(
+ f"freq_mask_smooth_hz needs to be at least {int((self.sr / (self._n_fft / 2)))} Hz"
+ )
+
+ n_grad_time = (
+ 1
+ if self.time_mask_smooth_ms is None
+ else int(self.time_mask_smooth_ms / ((self.hop_length / self.sr) * 1000))
+ )
+ if n_grad_time < 1:
+ raise ValueError(
+ f"time_mask_smooth_ms needs to be at least {int((self.hop_length / self.sr) * 1000)} ms"
+ )
+
+ if n_grad_time == 1 and n_grad_freq == 1:
+ return None
+
+ v_f = torch.cat(
+ [
+ linspace(0, 1, n_grad_freq + 1, endpoint=False),
+ linspace(1, 0, n_grad_freq + 2),
+ ]
+ )[1:-1]
+ v_t = torch.cat(
+ [
+ linspace(0, 1, n_grad_time + 1, endpoint=False),
+ linspace(1, 0, n_grad_time + 2),
+ ]
+ )[1:-1]
+ smoothing_filter = torch.outer(v_f, v_t).unsqueeze(0).unsqueeze(0)
+
+ return smoothing_filter / smoothing_filter.sum()
+
+ @torch.no_grad()
+ def _stationary_mask(
+ self, X_db: torch.Tensor, xn: Optional[torch.Tensor] = None
+ ) -> torch.Tensor:
+ """
+ Computes a stationary binary mask to filter out noise in a log-magnitude spectrogram.
+
+ Arguments:
+ X_db (torch.Tensor): 2D tensor of shape (frames, freq_bins) containing the log-magnitude spectrogram.
+ xn (torch.Tensor): 1D tensor containing the audio signal corresponding to X_db.
+
+ Returns:
+ sig_mask (torch.Tensor): Binary mask of the same shape as X_db, where values greater than the threshold
+ are set to 1, and the rest are set to 0.
+ """
+ if xn is not None:
+ if "privateuseone" in str(xn.device):
+ if not hasattr(self, "stft"):
+ self.stft = STFT(
+ filter_length=self.n_fft,
+ hop_length=self.hop_length,
+ win_length=self.win_length,
+ window="hann",
+ ).to(xn.device)
+ XN = self.stft.transform(xn)
+ else:
+ XN = torch.stft(
+ xn,
+ n_fft=self.n_fft,
+ hop_length=self.hop_length,
+ win_length=self.win_length,
+ return_complex=True,
+ pad_mode="constant",
+ center=True,
+ window=torch.hann_window(self.win_length).to(xn.device),
+ )
+ XN_db = amp_to_db(XN).to(dtype=X_db.dtype)
+ else:
+ XN_db = X_db
+
+ # calculate mean and standard deviation along the frequency axis
+ std_freq_noise, mean_freq_noise = torch.std_mean(XN_db, dim=-1)
+
+ # compute noise threshold
+ noise_thresh = mean_freq_noise + std_freq_noise * self.n_std_thresh_stationary
+
+ # create binary mask by thresholding the spectrogram
+ sig_mask = X_db > noise_thresh.unsqueeze(2)
+ return sig_mask
+
+ @torch.no_grad()
+ def _nonstationary_mask(self, X_abs: torch.Tensor) -> torch.Tensor:
+ """
+ Computes a non-stationary binary mask to filter out noise in a log-magnitude spectrogram.
+
+ Arguments:
+ X_abs (torch.Tensor): 2D tensor of shape (frames, freq_bins) containing the magnitude spectrogram.
+
+ Returns:
+ sig_mask (torch.Tensor): Binary mask of the same shape as X_abs, where values greater than the threshold
+ are set to 1, and the rest are set to 0.
+ """
+ X_smoothed = (
+ conv1d(
+ X_abs.reshape(-1, 1, X_abs.shape[-1]),
+ torch.ones(
+ self.n_movemean_nonstationary,
+ dtype=X_abs.dtype,
+ device=X_abs.device,
+ ).view(1, 1, -1),
+ padding="same",
+ ).view(X_abs.shape)
+ / self.n_movemean_nonstationary
+ )
+
+ # Compute slowness ratio and apply temperature sigmoid
+ slowness_ratio = (X_abs - X_smoothed) / (X_smoothed + 1e-6)
+ sig_mask = temperature_sigmoid(
+ slowness_ratio, self.n_thresh_nonstationary, self.temp_coeff_nonstationary
+ )
+
+ return sig_mask
+
+ def forward(
+ self, x: torch.Tensor, xn: Optional[torch.Tensor] = None
+ ) -> torch.Tensor:
+ """
+ Apply the proposed algorithm to the input signal.
+
+ Arguments:
+ x (torch.Tensor): The input audio signal, with shape (batch_size, signal_length).
+ xn (Optional[torch.Tensor]): The noise signal used for stationary noise reduction. If `None`, the input
+ signal is used as the noise signal. Default: `None`.
+
+ Returns:
+ torch.Tensor: The denoised audio signal, with the same shape as the input signal.
+ """
+
+ # Compute short-time Fourier transform (STFT)
+ if "privateuseone" in str(x.device):
+ if not hasattr(self, "stft"):
+ self.stft = STFT(
+ filter_length=self.n_fft,
+ hop_length=self.hop_length,
+ win_length=self.win_length,
+ window="hann",
+ ).to(x.device)
+ X, phase = self.stft.transform(x, return_phase=True)
+ else:
+ X = torch.stft(
+ x,
+ n_fft=self.n_fft,
+ hop_length=self.hop_length,
+ win_length=self.win_length,
+ return_complex=True,
+ pad_mode="constant",
+ center=True,
+ window=torch.hann_window(self.win_length).to(x.device),
+ )
+
+ # Compute signal mask based on stationary or nonstationary assumptions
+ if self.nonstationary:
+ sig_mask = self._nonstationary_mask(X.abs())
+ else:
+ sig_mask = self._stationary_mask(amp_to_db(X), xn)
+
+ # Propagate decrease in signal power
+ sig_mask = self.prop_decrease * (sig_mask.float() - 1.0) + 1.0
+
+ # Smooth signal mask with 2D convolution
+ if self.smoothing_filter is not None:
+ sig_mask = conv2d(
+ sig_mask.unsqueeze(1),
+ self.smoothing_filter.to(sig_mask.dtype),
+ padding="same",
+ )
+
+ # Apply signal mask to STFT magnitude and phase components
+ Y = X * sig_mask.squeeze(1)
+
+ # Inverse STFT to obtain time-domain signal
+ if "privateuseone" in str(Y.device):
+ y = self.stft.inverse(Y, phase)
+ else:
+ y = torch.istft(
+ Y,
+ n_fft=self.n_fft,
+ hop_length=self.hop_length,
+ win_length=self.win_length,
+ center=True,
+ window=torch.hann_window(self.win_length).to(Y.device),
+ )
+
+ return y.to(dtype=x.dtype)
diff --git a/infer/modules/gui/utils.py b/infer/modules/gui/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..46820981c9f60fd2200b2418b2b9e9f2e4d1fb10
--- /dev/null
+++ b/infer/modules/gui/utils.py
@@ -0,0 +1,70 @@
+import torch
+from torch.types import Number
+
+
+@torch.no_grad()
+def amp_to_db(
+ x: torch.Tensor, eps=torch.finfo(torch.float64).eps, top_db=40
+) -> torch.Tensor:
+ """
+ Convert the input tensor from amplitude to decibel scale.
+
+ Arguments:
+ x {[torch.Tensor]} -- [Input tensor.]
+
+ Keyword Arguments:
+ eps {[float]} -- [Small value to avoid numerical instability.]
+ (default: {torch.finfo(torch.float64).eps})
+ top_db {[float]} -- [threshold the output at ``top_db`` below the peak]
+ ` (default: {40})
+
+ Returns:
+ [torch.Tensor] -- [Output tensor in decibel scale.]
+ """
+ x_db = 20 * torch.log10(x.abs() + eps)
+ return torch.max(x_db, (x_db.max(-1).values - top_db).unsqueeze(-1))
+
+
+@torch.no_grad()
+def temperature_sigmoid(x: torch.Tensor, x0: float, temp_coeff: float) -> torch.Tensor:
+ """
+ Apply a sigmoid function with temperature scaling.
+
+ Arguments:
+ x {[torch.Tensor]} -- [Input tensor.]
+ x0 {[float]} -- [Parameter that controls the threshold of the sigmoid.]
+ temp_coeff {[float]} -- [Parameter that controls the slope of the sigmoid.]
+
+ Returns:
+ [torch.Tensor] -- [Output tensor after applying the sigmoid with temperature scaling.]
+ """
+ return torch.sigmoid((x - x0) / temp_coeff)
+
+
+@torch.no_grad()
+def linspace(
+ start: Number, stop: Number, num: int = 50, endpoint: bool = True, **kwargs
+) -> torch.Tensor:
+ """
+ Generate a linearly spaced 1-D tensor.
+
+ Arguments:
+ start {[Number]} -- [The starting value of the sequence.]
+ stop {[Number]} -- [The end value of the sequence, unless `endpoint` is set to False.
+ In that case, the sequence consists of all but the last of ``num + 1``
+ evenly spaced samples, so that `stop` is excluded. Note that the step
+ size changes when `endpoint` is False.]
+
+ Keyword Arguments:
+ num {[int]} -- [Number of samples to generate. Default is 50. Must be non-negative.]
+ endpoint {[bool]} -- [If True, `stop` is the last sample. Otherwise, it is not included.
+ Default is True.]
+ **kwargs -- [Additional arguments to be passed to the underlying PyTorch `linspace` function.]
+
+ Returns:
+ [torch.Tensor] -- [1-D tensor of `num` equally spaced samples from `start` to `stop`.]
+ """
+ if endpoint:
+ return torch.linspace(start, stop, num, **kwargs)
+ else:
+ return torch.linspace(start, stop, num + 1, **kwargs)[:-1]
diff --git a/infer/modules/train/extract/extract_f0_print.py b/infer/modules/train/extract/extract_f0_print.py
new file mode 100644
index 0000000000000000000000000000000000000000..230969218edf11354516df6c6673391ac7c1722b
--- /dev/null
+++ b/infer/modules/train/extract/extract_f0_print.py
@@ -0,0 +1,175 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+from multiprocessing import Process
+
+exp_dir = sys.argv[1]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+ print(strr)
+ f.write("%s\n" % strr)
+ f.flush()
+
+
+n_p = int(sys.argv[2])
+f0method = sys.argv[3]
+
+
+class FeatureInput(object):
+ def __init__(self, samplerate=16000, hop_size=160):
+ self.fs = samplerate
+ self.hop = hop_size
+
+ self.f0_bin = 256
+ self.f0_max = 1100.0
+ self.f0_min = 50.0
+ self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+ self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+ def compute_f0(self, path, f0_method):
+ x = load_audio(path, self.fs)
+ p_len = x.shape[0] // self.hop
+ if f0_method == "pm":
+ time_step = 160 / 16000 * 1000
+ f0_min = 50
+ f0_max = 1100
+ f0 = (
+ parselmouth.Sound(x, self.fs)
+ .to_pitch_ac(
+ time_step=time_step / 1000,
+ voicing_threshold=0.6,
+ pitch_floor=f0_min,
+ pitch_ceiling=f0_max,
+ )
+ .selected_array["frequency"]
+ )
+ pad_size = (p_len - len(f0) + 1) // 2
+ if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+ f0 = np.pad(
+ f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant"
+ )
+ elif f0_method == "harvest":
+ f0, t = pyworld.harvest(
+ x.astype(np.double),
+ fs=self.fs,
+ f0_ceil=self.f0_max,
+ f0_floor=self.f0_min,
+ frame_period=1000 * self.hop / self.fs,
+ )
+ f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+ elif f0_method == "dio":
+ f0, t = pyworld.dio(
+ x.astype(np.double),
+ fs=self.fs,
+ f0_ceil=self.f0_max,
+ f0_floor=self.f0_min,
+ frame_period=1000 * self.hop / self.fs,
+ )
+ f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.fs)
+ elif f0_method == "rmvpe":
+ if hasattr(self, "model_rmvpe") == False:
+ from rvc.f0.rmvpe import RMVPE
+
+ print("Loading rmvpe model")
+ self.model_rmvpe = RMVPE(
+ "assets/rmvpe/rmvpe.pt", is_half=False, device="cpu"
+ )
+ f0 = self.model_rmvpe.compute_f0(x, filter_radius=0.03)
+ return f0
+
+ def coarse_f0(self, f0):
+ f0_mel = 1127 * np.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+ self.f0_bin - 2
+ ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+ # use 0 or 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+ f0_coarse = np.rint(f0_mel).astype(int)
+ assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+ f0_coarse.max(),
+ f0_coarse.min(),
+ )
+ return f0_coarse
+
+ def go(self, paths, f0_method):
+ if len(paths) == 0:
+ printt("no-f0-todo")
+ else:
+ printt("todo-f0-%s" % len(paths))
+ n = max(len(paths) // 5, 1) # 每个进程最多打印5条
+ for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+ try:
+ if idx % n == 0:
+ printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+ if (
+ os.path.exists(opt_path1 + ".npy") == True
+ and os.path.exists(opt_path2 + ".npy") == True
+ ):
+ continue
+ featur_pit = self.compute_f0(inp_path, f0_method)
+ np.save(
+ opt_path2,
+ featur_pit,
+ allow_pickle=False,
+ ) # nsf
+ coarse_pit = self.coarse_f0(featur_pit)
+ np.save(
+ opt_path1,
+ coarse_pit,
+ allow_pickle=False,
+ ) # ori
+ except:
+ printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+ # exp_dir=r"E:\codes\py39\dataset\mi-test"
+ # n_p=16
+ # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+ printt(" ".join(sys.argv))
+ featureInput = FeatureInput()
+ paths = []
+ inp_root = "%s/1_16k_wavs" % (exp_dir)
+ opt_root1 = "%s/2a_f0" % (exp_dir)
+ opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+ os.makedirs(opt_root1, exist_ok=True)
+ os.makedirs(opt_root2, exist_ok=True)
+ for name in sorted(list(os.listdir(inp_root))):
+ inp_path = "%s/%s" % (inp_root, name)
+ if "spec" in inp_path:
+ continue
+ opt_path1 = "%s/%s" % (opt_root1, name)
+ opt_path2 = "%s/%s" % (opt_root2, name)
+ paths.append([inp_path, opt_path1, opt_path2])
+
+ ps = []
+ for i in range(n_p):
+ p = Process(
+ target=featureInput.go,
+ args=(
+ paths[i::n_p],
+ f0method,
+ ),
+ )
+ ps.append(p)
+ p.start()
+ for i in range(n_p):
+ ps[i].join()
diff --git a/infer/modules/train/extract/extract_f0_rmvpe.py b/infer/modules/train/extract/extract_f0_rmvpe.py
new file mode 100644
index 0000000000000000000000000000000000000000..68850fd948d680e5481bab7e56d8ee09282a290f
--- /dev/null
+++ b/infer/modules/train/extract/extract_f0_rmvpe.py
@@ -0,0 +1,141 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+n_part = int(sys.argv[1])
+i_part = int(sys.argv[2])
+i_gpu = sys.argv[3]
+os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+exp_dir = sys.argv[4]
+is_half = sys.argv[5]
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+ print(strr)
+ f.write("%s\n" % strr)
+ f.flush()
+
+
+class FeatureInput(object):
+ def __init__(self, samplerate=16000, hop_size=160):
+ self.fs = samplerate
+ self.hop = hop_size
+
+ self.f0_bin = 256
+ self.f0_max = 1100.0
+ self.f0_min = 50.0
+ self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+ self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+ def compute_f0(self, path, f0_method):
+ x = load_audio(path, self.fs)
+ # p_len = x.shape[0] // self.hop
+ if f0_method == "rmvpe":
+ if hasattr(self, "model_rmvpe") == False:
+ from rvc.f0.rmvpe import RMVPE
+
+ print("Loading rmvpe model")
+ self.model_rmvpe = RMVPE(
+ "assets/rmvpe/rmvpe.pt", is_half=is_half, device="cuda"
+ )
+ f0 = self.model_rmvpe.compute_f0(x, filter_radius=0.03)
+ return f0
+
+ def coarse_f0(self, f0):
+ f0_mel = 1127 * np.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+ self.f0_bin - 2
+ ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+ # use 0 or 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+ f0_coarse = np.rint(f0_mel).astype(int)
+ assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+ f0_coarse.max(),
+ f0_coarse.min(),
+ )
+ return f0_coarse
+
+ def go(self, paths, f0_method):
+ if len(paths) == 0:
+ printt("no-f0-todo")
+ else:
+ printt("todo-f0-%s" % len(paths))
+ n = max(len(paths) // 5, 1) # 每个进程最多打印5条
+ for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+ try:
+ if idx % n == 0:
+ printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+ if (
+ os.path.exists(opt_path1 + ".npy") == True
+ and os.path.exists(opt_path2 + ".npy") == True
+ ):
+ continue
+ featur_pit = self.compute_f0(inp_path, f0_method)
+ np.save(
+ opt_path2,
+ featur_pit,
+ allow_pickle=False,
+ ) # nsf
+ coarse_pit = self.coarse_f0(featur_pit)
+ np.save(
+ opt_path1,
+ coarse_pit,
+ allow_pickle=False,
+ ) # ori
+ except:
+ printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+ # exp_dir=r"E:\codes\py39\dataset\mi-test"
+ # n_p=16
+ # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+ printt(" ".join(sys.argv))
+ featureInput = FeatureInput()
+ paths = []
+ inp_root = "%s/1_16k_wavs" % (exp_dir)
+ opt_root1 = "%s/2a_f0" % (exp_dir)
+ opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+ os.makedirs(opt_root1, exist_ok=True)
+ os.makedirs(opt_root2, exist_ok=True)
+ for name in sorted(list(os.listdir(inp_root))):
+ inp_path = "%s/%s" % (inp_root, name)
+ if "spec" in inp_path:
+ continue
+ opt_path1 = "%s/%s" % (opt_root1, name)
+ opt_path2 = "%s/%s" % (opt_root2, name)
+ paths.append([inp_path, opt_path1, opt_path2])
+ try:
+ featureInput.go(paths[i_part::n_part], "rmvpe")
+ except:
+ printt("f0_all_fail-%s" % (traceback.format_exc()))
+ # ps = []
+ # for i in range(n_p):
+ # p = Process(
+ # target=featureInput.go,
+ # args=(
+ # paths[i::n_p],
+ # f0method,
+ # ),
+ # )
+ # ps.append(p)
+ # p.start()
+ # for i in range(n_p):
+ # ps[i].join()
diff --git a/infer/modules/train/extract/extract_f0_rmvpe_dml.py b/infer/modules/train/extract/extract_f0_rmvpe_dml.py
new file mode 100644
index 0000000000000000000000000000000000000000..404d7ef9f83bf584790a415e9b23c875e8e9ca55
--- /dev/null
+++ b/infer/modules/train/extract/extract_f0_rmvpe_dml.py
@@ -0,0 +1,139 @@
+import os
+import sys
+import traceback
+
+import parselmouth
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import logging
+
+import numpy as np
+import pyworld
+
+from infer.lib.audio import load_audio
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+exp_dir = sys.argv[1]
+import torch_directml
+
+device = torch_directml.device(torch_directml.default_device())
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+ print(strr)
+ f.write("%s\n" % strr)
+ f.flush()
+
+
+class FeatureInput(object):
+ def __init__(self, samplerate=16000, hop_size=160):
+ self.fs = samplerate
+ self.hop = hop_size
+
+ self.f0_bin = 256
+ self.f0_max = 1100.0
+ self.f0_min = 50.0
+ self.f0_mel_min = 1127 * np.log(1 + self.f0_min / 700)
+ self.f0_mel_max = 1127 * np.log(1 + self.f0_max / 700)
+
+ def compute_f0(self, path, f0_method):
+ x = load_audio(path, self.fs)
+ # p_len = x.shape[0] // self.hop
+ if f0_method == "rmvpe":
+ if hasattr(self, "model_rmvpe") == False:
+ from rvc.f0.rmvpe import RMVPE
+
+ print("Loading rmvpe model")
+ self.model_rmvpe = RMVPE(
+ "assets/rmvpe/rmvpe.pt", is_half=False, device=device
+ )
+ f0 = self.model_rmvpe.compute_f0(x, filter_radius=0.03)
+ return f0
+
+ def coarse_f0(self, f0):
+ f0_mel = 1127 * np.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - self.f0_mel_min) * (
+ self.f0_bin - 2
+ ) / (self.f0_mel_max - self.f0_mel_min) + 1
+
+ # use 0 or 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > self.f0_bin - 1] = self.f0_bin - 1
+ f0_coarse = np.rint(f0_mel).astype(int)
+ assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
+ f0_coarse.max(),
+ f0_coarse.min(),
+ )
+ return f0_coarse
+
+ def go(self, paths, f0_method):
+ if len(paths) == 0:
+ printt("no-f0-todo")
+ else:
+ printt("todo-f0-%s" % len(paths))
+ n = max(len(paths) // 5, 1) # 每个进程最多打印5条
+ for idx, (inp_path, opt_path1, opt_path2) in enumerate(paths):
+ try:
+ if idx % n == 0:
+ printt("f0ing,now-%s,all-%s,-%s" % (idx, len(paths), inp_path))
+ if (
+ os.path.exists(opt_path1 + ".npy") == True
+ and os.path.exists(opt_path2 + ".npy") == True
+ ):
+ continue
+ featur_pit = self.compute_f0(inp_path, f0_method)
+ np.save(
+ opt_path2,
+ featur_pit,
+ allow_pickle=False,
+ ) # nsf
+ coarse_pit = self.coarse_f0(featur_pit)
+ np.save(
+ opt_path1,
+ coarse_pit,
+ allow_pickle=False,
+ ) # ori
+ except:
+ printt("f0fail-%s-%s-%s" % (idx, inp_path, traceback.format_exc()))
+
+
+if __name__ == "__main__":
+ # exp_dir=r"E:\codes\py39\dataset\mi-test"
+ # n_p=16
+ # f = open("%s/log_extract_f0.log"%exp_dir, "w")
+ printt(" ".join(sys.argv))
+ featureInput = FeatureInput()
+ paths = []
+ inp_root = "%s/1_16k_wavs" % (exp_dir)
+ opt_root1 = "%s/2a_f0" % (exp_dir)
+ opt_root2 = "%s/2b-f0nsf" % (exp_dir)
+
+ os.makedirs(opt_root1, exist_ok=True)
+ os.makedirs(opt_root2, exist_ok=True)
+ for name in sorted(list(os.listdir(inp_root))):
+ inp_path = "%s/%s" % (inp_root, name)
+ if "spec" in inp_path:
+ continue
+ opt_path1 = "%s/%s" % (opt_root1, name)
+ opt_path2 = "%s/%s" % (opt_root2, name)
+ paths.append([inp_path, opt_path1, opt_path2])
+ try:
+ featureInput.go(paths, "rmvpe")
+ except:
+ printt("f0_all_fail-%s" % (traceback.format_exc()))
+ # ps = []
+ # for i in range(n_p):
+ # p = Process(
+ # target=featureInput.go,
+ # args=(
+ # paths[i::n_p],
+ # f0method,
+ # ),
+ # )
+ # ps.append(p)
+ # p.start()
+ # for i in range(n_p):
+ # ps[i].join()
diff --git a/infer/modules/train/extract_feature_print.py b/infer/modules/train/extract_feature_print.py
new file mode 100644
index 0000000000000000000000000000000000000000..96a69dee4614dc9c7cbafd24247f244b4e28f9d4
--- /dev/null
+++ b/infer/modules/train/extract_feature_print.py
@@ -0,0 +1,142 @@
+import os
+import sys
+import traceback
+
+os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+os.environ["PYTORCH_MPS_HIGH_WATERMARK_RATIO"] = "0.0"
+
+device = sys.argv[1]
+n_part = int(sys.argv[2])
+i_part = int(sys.argv[3])
+if len(sys.argv) == 7:
+ exp_dir = sys.argv[4]
+ version = sys.argv[5]
+ is_half = sys.argv[6].lower() == "true"
+else:
+ i_gpu = sys.argv[4]
+ exp_dir = sys.argv[5]
+ os.environ["CUDA_VISIBLE_DEVICES"] = str(i_gpu)
+ version = sys.argv[6]
+ is_half = sys.argv[7].lower() == "true"
+import fairseq
+import numpy as np
+import soundfile as sf
+import torch
+import torch.nn.functional as F
+
+if "privateuseone" not in device:
+ device = "cpu"
+ if torch.cuda.is_available():
+ device = "cuda"
+ elif torch.backends.mps.is_available():
+ device = "mps"
+else:
+ import torch_directml
+
+ device = torch_directml.device(torch_directml.default_device())
+
+ def forward_dml(ctx, x, scale):
+ ctx.scale = scale
+ res = x.clone().detach()
+ return res
+
+ fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+
+f = open("%s/extract_f0_feature.log" % exp_dir, "a+")
+
+
+def printt(strr):
+ print(strr)
+ f.write("%s\n" % strr)
+ f.flush()
+
+
+printt(" ".join(sys.argv))
+model_path = "assets/hubert/hubert_base.pt"
+
+printt("exp_dir: " + exp_dir)
+wavPath = "%s/1_16k_wavs" % exp_dir
+outPath = (
+ "%s/3_feature256" % exp_dir if version == "v1" else "%s/3_feature768" % exp_dir
+)
+os.makedirs(outPath, exist_ok=True)
+
+
+# wave must be 16k, hop_size=320
+def readwave(wav_path, normalize=False):
+ wav, sr = sf.read(wav_path)
+ assert sr == 16000
+ feats = torch.from_numpy(wav).float()
+ if feats.dim() == 2: # double channels
+ feats = feats.mean(-1)
+ assert feats.dim() == 1, feats.dim()
+ if normalize:
+ with torch.no_grad():
+ feats = F.layer_norm(feats, feats.shape)
+ feats = feats.view(1, -1)
+ return feats
+
+
+# HuBERT model
+printt("load model(s) from {}".format(model_path))
+# if hubert model is exist
+if os.access(model_path, os.F_OK) == False:
+ printt(
+ "Error: Extracting is shut down because %s does not exist, you may download it from https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main"
+ % model_path
+ )
+ exit(0)
+models, saved_cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task(
+ [model_path],
+ suffix="",
+)
+model = models[0]
+model = model.to(device)
+printt("move model to %s" % device)
+if is_half:
+ if device not in ["mps", "cpu"]:
+ model = model.half()
+model.eval()
+
+todo = sorted(list(os.listdir(wavPath)))[i_part::n_part]
+n = max(1, len(todo) // 10) # 最多打印十条
+if len(todo) == 0:
+ printt("no-feature-todo")
+else:
+ printt("all-feature-%s" % len(todo))
+ for idx, file in enumerate(todo):
+ try:
+ if file.endswith(".wav"):
+ wav_path = "%s/%s" % (wavPath, file)
+ out_path = "%s/%s" % (outPath, file.replace("wav", "npy"))
+
+ if os.path.exists(out_path):
+ continue
+
+ feats = readwave(wav_path, normalize=saved_cfg.task.normalize)
+ padding_mask = torch.BoolTensor(feats.shape).fill_(False)
+ inputs = {
+ "source": (
+ feats.half().to(device)
+ if is_half and device not in ["mps", "cpu"]
+ else feats.to(device)
+ ),
+ "padding_mask": padding_mask.to(device),
+ "output_layer": 9 if version == "v1" else 12, # layer 9
+ }
+ with torch.no_grad():
+ logits = model.extract_features(**inputs)
+ feats = (
+ model.final_proj(logits[0]) if version == "v1" else logits[0]
+ )
+
+ feats = feats.squeeze(0).float().cpu().numpy()
+ if np.isnan(feats).sum() == 0:
+ np.save(out_path, feats, allow_pickle=False)
+ else:
+ printt("%s-contains nan" % file)
+ if idx % n == 0:
+ printt("now-%s,all-%s,%s,%s" % (len(todo), idx, file, feats.shape))
+ except:
+ printt(traceback.format_exc())
+ printt("all-feature-done")
diff --git a/infer/modules/train/preprocess.py b/infer/modules/train/preprocess.py
new file mode 100644
index 0000000000000000000000000000000000000000..138bb4c642ce0373b4718ea623768b4460245a6e
--- /dev/null
+++ b/infer/modules/train/preprocess.py
@@ -0,0 +1,142 @@
+import multiprocessing
+import os
+import sys
+
+from scipy import signal
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+print(*sys.argv[1:])
+inp_root = sys.argv[1]
+sr = int(sys.argv[2])
+n_p = int(sys.argv[3])
+exp_dir = sys.argv[4]
+noparallel = sys.argv[5] == "True"
+per = float(sys.argv[6])
+import os
+import traceback
+
+import librosa
+import numpy as np
+from scipy.io import wavfile
+
+from infer.lib.audio import load_audio
+from infer.lib.slicer2 import Slicer
+
+f = open("%s/preprocess.log" % exp_dir, "a+")
+
+
+def println(strr):
+ print(strr)
+ f.write("%s\n" % strr)
+ f.flush()
+
+
+class PreProcess:
+ def __init__(self, sr, exp_dir, per=3.7):
+ self.slicer = Slicer(
+ sr=sr,
+ threshold=-42,
+ min_length=1500,
+ min_interval=400,
+ hop_size=15,
+ max_sil_kept=500,
+ )
+ self.sr = sr
+ self.bh, self.ah = signal.butter(N=5, Wn=48, btype="high", fs=self.sr)
+ self.per = per
+ self.overlap = 0.3
+ self.tail = self.per + self.overlap
+ self.max = 0.9
+ self.alpha = 0.75
+ self.exp_dir = exp_dir
+ self.gt_wavs_dir = "%s/0_gt_wavs" % exp_dir
+ self.wavs16k_dir = "%s/1_16k_wavs" % exp_dir
+ os.makedirs(self.exp_dir, exist_ok=True)
+ os.makedirs(self.gt_wavs_dir, exist_ok=True)
+ os.makedirs(self.wavs16k_dir, exist_ok=True)
+
+ def norm_write(self, tmp_audio, idx0, idx1):
+ tmp_max = np.abs(tmp_audio).max()
+ if tmp_max > 2.5:
+ print("%s-%s-%s-filtered" % (idx0, idx1, tmp_max))
+ return
+ tmp_audio = (tmp_audio / tmp_max * (self.max * self.alpha)) + (
+ 1 - self.alpha
+ ) * tmp_audio
+ wavfile.write(
+ "%s/%s_%s.wav" % (self.gt_wavs_dir, idx0, idx1),
+ self.sr,
+ tmp_audio.astype(np.float32),
+ )
+ tmp_audio = librosa.resample(
+ tmp_audio, orig_sr=self.sr, target_sr=16000
+ ) # , res_type="soxr_vhq"
+ wavfile.write(
+ "%s/%s_%s.wav" % (self.wavs16k_dir, idx0, idx1),
+ 16000,
+ tmp_audio.astype(np.float32),
+ )
+
+ def pipeline(self, path, idx0):
+ try:
+ audio = load_audio(path, self.sr)
+ # zero phased digital filter cause pre-ringing noise...
+ # audio = signal.filtfilt(self.bh, self.ah, audio)
+ audio = signal.lfilter(self.bh, self.ah, audio)
+
+ idx1 = 0
+ for audio in self.slicer.slice(audio):
+ i = 0
+ while 1:
+ start = int(self.sr * (self.per - self.overlap) * i)
+ i += 1
+ if len(audio[start:]) > self.tail * self.sr:
+ tmp_audio = audio[start : start + int(self.per * self.sr)]
+ self.norm_write(tmp_audio, idx0, idx1)
+ idx1 += 1
+ else:
+ tmp_audio = audio[start:]
+ idx1 += 1
+ break
+ self.norm_write(tmp_audio, idx0, idx1)
+ println("%s\t-> Success" % path)
+ except:
+ println("%s\t-> %s" % (path, traceback.format_exc()))
+
+ def pipeline_mp(self, infos):
+ for path, idx0 in infos:
+ self.pipeline(path, idx0)
+
+ def pipeline_mp_inp_dir(self, inp_root, n_p):
+ try:
+ infos = [
+ ("%s/%s" % (inp_root, name), idx)
+ for idx, name in enumerate(sorted(list(os.listdir(inp_root))))
+ ]
+ if noparallel:
+ for i in range(n_p):
+ self.pipeline_mp(infos[i::n_p])
+ else:
+ ps = []
+ for i in range(n_p):
+ p = multiprocessing.Process(
+ target=self.pipeline_mp, args=(infos[i::n_p],)
+ )
+ ps.append(p)
+ p.start()
+ for i in range(n_p):
+ ps[i].join()
+ except:
+ println("Fail. %s" % traceback.format_exc())
+
+
+def preprocess_trainset(inp_root, sr, n_p, exp_dir, per):
+ pp = PreProcess(sr, exp_dir, per)
+ println("start preprocess")
+ pp.pipeline_mp_inp_dir(inp_root, n_p)
+ println("end preprocess")
+
+
+if __name__ == "__main__":
+ preprocess_trainset(inp_root, sr, n_p, exp_dir, per)
diff --git a/infer/modules/train/train.py b/infer/modules/train/train.py
new file mode 100644
index 0000000000000000000000000000000000000000..8d32e4a7c5703e8e53c59b153c9c6181fbef8e07
--- /dev/null
+++ b/infer/modules/train/train.py
@@ -0,0 +1,652 @@
+import os
+import sys
+import logging
+from typing import Tuple
+
+logger = logging.getLogger(__name__)
+logging.getLogger("numba").setLevel(logging.WARNING)
+
+now_dir = os.getcwd()
+sys.path.append(os.path.join(now_dir))
+
+import datetime
+
+from infer.lib.train import utils
+
+hps = utils.get_hparams()
+os.environ["CUDA_VISIBLE_DEVICES"] = hps.gpus.replace("-", ",")
+n_gpus = len(hps.gpus.split("-"))
+from random import randint, shuffle
+
+import torch
+
+try:
+ import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+
+ if torch.xpu.is_available():
+ from rvc.ipex import ipex_init, gradscaler_init
+ from torch.xpu.amp import autocast
+
+ GradScaler = gradscaler_init()
+ ipex_init()
+ else:
+ from torch.cuda.amp import GradScaler, autocast
+except Exception:
+ from torch.cuda.amp import GradScaler, autocast
+
+torch.backends.cudnn.deterministic = False
+torch.backends.cudnn.benchmark = False
+from time import sleep
+from time import time as ttime
+
+import torch.distributed as dist
+import torch.multiprocessing as mp
+from torch.nn import functional as F
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.data import DataLoader
+from torch.utils.tensorboard import SummaryWriter
+
+from infer.lib.train.data_utils import (
+ DistributedBucketSampler,
+ TextAudioCollate,
+ TextAudioCollateMultiNSFsid,
+ TextAudioLoader,
+ TextAudioLoaderMultiNSFsid,
+)
+
+from rvc.layers.discriminators import MultiPeriodDiscriminator
+
+if hps.version == "v1":
+ from rvc.layers.synthesizers import SynthesizerTrnMs256NSFsid as RVC_Model_f0
+ from rvc.layers.synthesizers import (
+ SynthesizerTrnMs256NSFsid_nono as RVC_Model_nof0,
+ )
+else:
+ from rvc.layers.synthesizers import (
+ SynthesizerTrnMs768NSFsid as RVC_Model_f0,
+ SynthesizerTrnMs768NSFsid_nono as RVC_Model_nof0,
+ )
+
+from infer.lib.train.losses import (
+ discriminator_loss,
+ feature_loss,
+ generator_loss,
+ kl_loss,
+)
+from infer.lib.train.mel_processing import mel_spectrogram_torch, spec_to_mel_torch
+from infer.lib.train.process_ckpt import save_small_model
+
+from rvc.layers.utils import (
+ slice_on_last_dim,
+ total_grad_norm,
+)
+
+global_step = 0
+
+
+class EpochRecorder:
+ def __init__(self):
+ self.last_time = ttime()
+
+ def record(self):
+ now_time = ttime()
+ elapsed_time = now_time - self.last_time
+ self.last_time = now_time
+ elapsed_time_str = str(datetime.timedelta(seconds=elapsed_time))
+ current_time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+ return f"[{current_time}] | ({elapsed_time_str})"
+
+
+def main():
+ n_gpus = torch.cuda.device_count()
+
+ if torch.cuda.is_available() == False and torch.backends.mps.is_available() == True:
+ n_gpus = 1
+ if n_gpus < 1:
+ # patch to unblock people without gpus. there is probably a better way.
+ print("NO GPU DETECTED: falling back to CPU - this may take a while")
+ n_gpus = 1
+ os.environ["MASTER_ADDR"] = "localhost"
+ os.environ["MASTER_PORT"] = str(randint(20000, 55555))
+ children = []
+ logger = utils.get_logger(hps.model_dir)
+ for i in range(n_gpus):
+ subproc = mp.Process(
+ target=run,
+ args=(i, n_gpus, hps, logger),
+ )
+ children.append(subproc)
+ subproc.start()
+
+ for i in range(n_gpus):
+ children[i].join()
+
+
+def run(rank, n_gpus, hps: utils.HParams, logger: logging.Logger):
+ global global_step
+ if rank == 0:
+ # logger = utils.get_logger(hps.model_dir)
+ logger.info(hps)
+ # utils.check_git_hash(hps.model_dir)
+ writer = SummaryWriter(log_dir=hps.model_dir)
+ writer_eval = SummaryWriter(log_dir=os.path.join(hps.model_dir, "eval"))
+
+ dist.init_process_group(
+ backend="gloo", init_method="env://", world_size=n_gpus, rank=rank
+ )
+ torch.manual_seed(hps.train.seed)
+ if torch.cuda.is_available():
+ torch.cuda.set_device(rank)
+
+ if hps.if_f0 == 1:
+ train_dataset = TextAudioLoaderMultiNSFsid(hps.data.training_files, hps.data)
+ else:
+ train_dataset = TextAudioLoader(hps.data.training_files, hps.data)
+ train_sampler = DistributedBucketSampler(
+ train_dataset,
+ hps.train.batch_size * n_gpus,
+ # [100, 200, 300, 400, 500, 600, 700, 800, 900, 1000, 1200,1400], # 16s
+ [100, 200, 300, 400, 500, 600, 700, 800, 900], # 16s
+ num_replicas=n_gpus,
+ rank=rank,
+ shuffle=True,
+ )
+ # It is possible that dataloader's workers are out of shared memory. Please try to raise your shared memory limit.
+ # num_workers=8 -> num_workers=4
+ if hps.if_f0 == 1:
+ collate_fn = TextAudioCollateMultiNSFsid()
+ else:
+ collate_fn = TextAudioCollate()
+ train_loader = DataLoader(
+ train_dataset,
+ num_workers=4,
+ shuffle=False,
+ pin_memory=True,
+ collate_fn=collate_fn,
+ batch_sampler=train_sampler,
+ persistent_workers=True,
+ prefetch_factor=8,
+ )
+ mdl = hps.copy().model
+ del mdl.use_spectral_norm
+ if hps.if_f0 == 1:
+ net_g = RVC_Model_f0(
+ hps.data.filter_length // 2 + 1,
+ hps.train.segment_size // hps.data.hop_length,
+ **mdl,
+ sr=hps.sample_rate,
+ )
+ else:
+ net_g = RVC_Model_nof0(
+ hps.data.filter_length // 2 + 1,
+ hps.train.segment_size // hps.data.hop_length,
+ **mdl,
+ )
+ if torch.cuda.is_available():
+ net_g = net_g.cuda(rank)
+ has_xpu = bool(hasattr(torch, "xpu") and torch.xpu.is_available())
+ net_d = MultiPeriodDiscriminator(
+ hps.version,
+ use_spectral_norm=hps.model.use_spectral_norm,
+ has_xpu=has_xpu,
+ )
+ if torch.cuda.is_available():
+ net_d = net_d.cuda(rank)
+ optim_g = torch.optim.AdamW(
+ net_g.parameters(),
+ hps.train.learning_rate,
+ betas=hps.train.betas,
+ eps=hps.train.eps,
+ )
+ optim_d = torch.optim.AdamW(
+ net_d.parameters(),
+ hps.train.learning_rate,
+ betas=hps.train.betas,
+ eps=hps.train.eps,
+ )
+ # net_g = DDP(net_g, device_ids=[rank], find_unused_parameters=True)
+ # net_d = DDP(net_d, device_ids=[rank], find_unused_parameters=True)
+ if hasattr(torch, "xpu") and torch.xpu.is_available():
+ pass
+ elif torch.cuda.is_available():
+ net_g = DDP(net_g, device_ids=[rank])
+ net_d = DDP(net_d, device_ids=[rank])
+ else:
+ net_g = DDP(net_g)
+ net_d = DDP(net_d)
+
+ try: # 如果能加载自动resume
+ _, _, _, epoch_str = utils.load_checkpoint(
+ utils.latest_checkpoint_path(hps.model_dir, "D_*.pth"), net_d, optim_d
+ ) # D多半加载没事
+ if rank == 0:
+ logger.info("loaded D")
+ # _, _, _, epoch_str = utils.load_checkpoint(utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g,load_opt=0)
+ _, _, _, epoch_str = utils.load_checkpoint(
+ utils.latest_checkpoint_path(hps.model_dir, "G_*.pth"), net_g, optim_g
+ )
+ global_step = (epoch_str - 1) * len(train_loader)
+ # epoch_str = 1
+ # global_step = 0
+ except: # 如果首次不能加载，加载pretrain
+ # traceback.print_exc()
+ epoch_str = 1
+ global_step = 0
+ if hps.pretrainG != "":
+ if rank == 0:
+ logger.info("loaded pretrained %s" % (hps.pretrainG))
+ if hasattr(net_g, "module"):
+ logger.info(
+ net_g.module.load_state_dict(
+ torch.load(hps.pretrainG, map_location="cpu")["model"]
+ )
+ ) ##测试不加载优化器
+ else:
+ logger.info(
+ net_g.load_state_dict(
+ torch.load(hps.pretrainG, map_location="cpu")["model"]
+ )
+ ) ##测试不加载优化器
+ if hps.pretrainD != "":
+ if rank == 0:
+ logger.info("loaded pretrained %s" % (hps.pretrainD))
+ if hasattr(net_d, "module"):
+ logger.info(
+ net_d.module.load_state_dict(
+ torch.load(hps.pretrainD, map_location="cpu")["model"]
+ )
+ )
+ else:
+ logger.info(
+ net_d.load_state_dict(
+ torch.load(hps.pretrainD, map_location="cpu")["model"]
+ )
+ )
+
+ scheduler_g = torch.optim.lr_scheduler.ExponentialLR(
+ optim_g, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
+ )
+ scheduler_d = torch.optim.lr_scheduler.ExponentialLR(
+ optim_d, gamma=hps.train.lr_decay, last_epoch=epoch_str - 2
+ )
+
+ scaler = GradScaler(enabled=hps.train.fp16_run)
+
+ cache = []
+ for epoch in range(epoch_str, hps.train.epochs + 1):
+ if rank == 0:
+ train_and_evaluate(
+ rank,
+ epoch,
+ hps,
+ [net_g, net_d],
+ [optim_g, optim_d],
+ [scheduler_g, scheduler_d],
+ scaler,
+ [train_loader, None],
+ logger,
+ [writer, writer_eval],
+ cache,
+ )
+ else:
+ train_and_evaluate(
+ rank,
+ epoch,
+ hps,
+ [net_g, net_d],
+ [optim_g, optim_d],
+ [scheduler_g, scheduler_d],
+ scaler,
+ [train_loader, None],
+ None,
+ None,
+ cache,
+ )
+ scheduler_g.step()
+ scheduler_d.step()
+
+
+def train_and_evaluate(
+ rank,
+ epoch,
+ hps,
+ nets: Tuple[RVC_Model_f0, MultiPeriodDiscriminator],
+ optims,
+ schedulers,
+ scaler,
+ loaders,
+ logger,
+ writers,
+ cache,
+):
+ net_g, net_d = nets
+ optim_g, optim_d = optims
+ train_loader, eval_loader = loaders
+ if writers is not None:
+ writer, writer_eval = writers
+
+ train_loader.batch_sampler.set_epoch(epoch)
+ global global_step
+
+ net_g.train()
+ net_d.train()
+
+ # Prepare data iterator
+ if hps.if_cache_data_in_gpu == True:
+ # Use Cache
+ data_iterator = cache
+ if cache == []:
+ # Make new cache
+ for batch_idx, info in enumerate(train_loader):
+ # Unpack
+ if hps.if_f0 == 1:
+ (
+ phone,
+ phone_lengths,
+ pitch,
+ pitchf,
+ spec,
+ spec_lengths,
+ wave,
+ wave_lengths,
+ sid,
+ ) = info
+ else:
+ (
+ phone,
+ phone_lengths,
+ spec,
+ spec_lengths,
+ wave,
+ wave_lengths,
+ sid,
+ ) = info
+ # Load on CUDA
+ if torch.cuda.is_available():
+ phone = phone.cuda(rank, non_blocking=True)
+ phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
+ if hps.if_f0 == 1:
+ pitch = pitch.cuda(rank, non_blocking=True)
+ pitchf = pitchf.cuda(rank, non_blocking=True)
+ sid = sid.cuda(rank, non_blocking=True)
+ spec = spec.cuda(rank, non_blocking=True)
+ spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
+ wave = wave.cuda(rank, non_blocking=True)
+ wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
+ # Cache on list
+ if hps.if_f0 == 1:
+ cache.append(
+ (
+ batch_idx,
+ (
+ phone,
+ phone_lengths,
+ pitch,
+ pitchf,
+ spec,
+ spec_lengths,
+ wave,
+ wave_lengths,
+ sid,
+ ),
+ )
+ )
+ else:
+ cache.append(
+ (
+ batch_idx,
+ (
+ phone,
+ phone_lengths,
+ spec,
+ spec_lengths,
+ wave,
+ wave_lengths,
+ sid,
+ ),
+ )
+ )
+ else:
+ # Load shuffled cache
+ shuffle(cache)
+ else:
+ # Loader
+ data_iterator = enumerate(train_loader)
+
+ # Run steps
+ epoch_recorder = EpochRecorder()
+ for batch_idx, info in data_iterator:
+ # Data
+ ## Unpack
+ pitch = pitchf = None
+ if hps.if_f0 == 1:
+ (
+ phone,
+ phone_lengths,
+ pitch,
+ pitchf,
+ spec,
+ spec_lengths,
+ wave,
+ wave_lengths,
+ sid,
+ ) = info
+ else:
+ phone, phone_lengths, spec, spec_lengths, wave, wave_lengths, sid = info
+ ## Load on CUDA
+ if (hps.if_cache_data_in_gpu == False) and torch.cuda.is_available():
+ phone = phone.cuda(rank, non_blocking=True)
+ phone_lengths = phone_lengths.cuda(rank, non_blocking=True)
+ if hps.if_f0 == 1:
+ pitch = pitch.cuda(rank, non_blocking=True)
+ pitchf = pitchf.cuda(rank, non_blocking=True)
+ sid = sid.cuda(rank, non_blocking=True)
+ spec = spec.cuda(rank, non_blocking=True)
+ spec_lengths = spec_lengths.cuda(rank, non_blocking=True)
+ wave = wave.cuda(rank, non_blocking=True)
+ # wave_lengths = wave_lengths.cuda(rank, non_blocking=True)
+
+ # Calculate
+ with autocast(enabled=hps.train.fp16_run):
+ (
+ y_hat,
+ ids_slice,
+ x_mask,
+ z_mask,
+ (z, z_p, m_p, logs_p, m_q, logs_q),
+ ) = net_g(phone, phone_lengths, spec, spec_lengths, sid, pitch, pitchf)
+ mel = spec_to_mel_torch(
+ spec,
+ hps.data.filter_length,
+ hps.data.n_mel_channels,
+ hps.data.sampling_rate,
+ hps.data.mel_fmin,
+ hps.data.mel_fmax,
+ )
+ y_mel = slice_on_last_dim(
+ mel, ids_slice, hps.train.segment_size // hps.data.hop_length
+ )
+ with autocast(enabled=False):
+ y_hat_mel = mel_spectrogram_torch(
+ y_hat.float().squeeze(1),
+ hps.data.filter_length,
+ hps.data.n_mel_channels,
+ hps.data.sampling_rate,
+ hps.data.hop_length,
+ hps.data.win_length,
+ hps.data.mel_fmin,
+ hps.data.mel_fmax,
+ )
+ if hps.train.fp16_run == True:
+ y_hat_mel = y_hat_mel.half()
+ wave = slice_on_last_dim(
+ wave, ids_slice * hps.data.hop_length, hps.train.segment_size
+ ) # slice
+
+ # Discriminator
+ y_d_hat_r, y_d_hat_g, _, _ = net_d(wave, y_hat.detach())
+ with autocast(enabled=False):
+ loss_disc, losses_disc_r, losses_disc_g = discriminator_loss(
+ y_d_hat_r, y_d_hat_g
+ )
+ optim_d.zero_grad()
+ scaler.scale(loss_disc).backward()
+ scaler.unscale_(optim_d)
+ grad_norm_d = total_grad_norm(net_d.parameters())
+ scaler.step(optim_d)
+
+ with autocast(enabled=hps.train.fp16_run):
+ # Generator
+ y_d_hat_r, y_d_hat_g, fmap_r, fmap_g = net_d(wave, y_hat)
+ with autocast(enabled=False):
+ loss_mel = F.l1_loss(y_mel, y_hat_mel) * hps.train.c_mel
+ loss_kl = kl_loss(z_p, logs_q, m_p, logs_p, z_mask) * hps.train.c_kl
+ loss_fm = feature_loss(fmap_r, fmap_g)
+ loss_gen, losses_gen = generator_loss(y_d_hat_g)
+ loss_gen_all = loss_gen + loss_fm + loss_mel + loss_kl
+ optim_g.zero_grad()
+ scaler.scale(loss_gen_all).backward()
+ scaler.unscale_(optim_g)
+ grad_norm_g = total_grad_norm(net_g.parameters())
+ scaler.step(optim_g)
+ scaler.update()
+
+ if rank == 0:
+ if global_step % hps.train.log_interval == 0:
+ lr = optim_g.param_groups[0]["lr"]
+ logger.info(
+ "Train Epoch: {} [{:.0f}%]".format(
+ epoch, 100.0 * batch_idx / len(train_loader)
+ )
+ )
+ # Amor For Tensorboard display
+ if loss_mel > 75:
+ loss_mel = 75
+ if loss_kl > 9:
+ loss_kl = 9
+
+ logger.info([global_step, lr])
+ logger.info(
+ f"loss_disc={loss_disc:.3f}, loss_gen={loss_gen:.3f}, loss_fm={loss_fm:.3f},loss_mel={loss_mel:.3f}, loss_kl={loss_kl:.3f}"
+ )
+ scalar_dict = {
+ "loss/g/total": loss_gen_all,
+ "loss/d/total": loss_disc,
+ "learning_rate": lr,
+ "grad_norm_d": grad_norm_d,
+ "grad_norm_g": grad_norm_g,
+ }
+ scalar_dict.update(
+ {
+ "loss/g/fm": loss_fm,
+ "loss/g/mel": loss_mel,
+ "loss/g/kl": loss_kl,
+ }
+ )
+
+ scalar_dict.update(
+ {"loss/g/{}".format(i): v for i, v in enumerate(losses_gen)}
+ )
+ scalar_dict.update(
+ {"loss/d_r/{}".format(i): v for i, v in enumerate(losses_disc_r)}
+ )
+ scalar_dict.update(
+ {"loss/d_g/{}".format(i): v for i, v in enumerate(losses_disc_g)}
+ )
+ image_dict = {
+ "slice/mel_org": utils.plot_spectrogram_to_numpy(
+ y_mel[0].data.cpu().numpy()
+ ),
+ "slice/mel_gen": utils.plot_spectrogram_to_numpy(
+ y_hat_mel[0].data.cpu().numpy()
+ ),
+ "all/mel": utils.plot_spectrogram_to_numpy(
+ mel[0].data.cpu().numpy()
+ ),
+ }
+ utils.summarize(
+ writer=writer,
+ global_step=global_step,
+ images=image_dict,
+ scalars=scalar_dict,
+ )
+ global_step += 1
+ # /Run steps
+
+ if epoch % hps.save_every_epoch == 0 and rank == 0:
+ if hps.if_latest == 0:
+ utils.save_checkpoint(
+ net_g,
+ optim_g,
+ hps.train.learning_rate,
+ epoch,
+ os.path.join(hps.model_dir, "G_{}.pth".format(global_step)),
+ )
+ utils.save_checkpoint(
+ net_d,
+ optim_d,
+ hps.train.learning_rate,
+ epoch,
+ os.path.join(hps.model_dir, "D_{}.pth".format(global_step)),
+ )
+ else:
+ utils.save_checkpoint(
+ net_g,
+ optim_g,
+ hps.train.learning_rate,
+ epoch,
+ os.path.join(hps.model_dir, "G_{}.pth".format(2333333)),
+ )
+ utils.save_checkpoint(
+ net_d,
+ optim_d,
+ hps.train.learning_rate,
+ epoch,
+ os.path.join(hps.model_dir, "D_{}.pth".format(2333333)),
+ )
+ if rank == 0 and hps.save_every_weights == "1":
+ if hasattr(net_g, "module"):
+ ckpt = net_g.module.state_dict()
+ else:
+ ckpt = net_g.state_dict()
+ logger.info(
+ "saving ckpt %s_e%s:%s"
+ % (
+ hps.name,
+ epoch,
+ save_small_model(
+ ckpt,
+ hps.sample_rate,
+ hps.if_f0,
+ hps.name + "_e%s_s%s" % (epoch, global_step),
+ epoch,
+ hps.version,
+ hps,
+ ),
+ )
+ )
+
+ if rank == 0:
+ logger.info("====> Epoch: {} {}".format(epoch, epoch_recorder.record()))
+ if epoch >= hps.total_epoch and rank == 0:
+ logger.info("Training is done. The program is closed.")
+
+ if hasattr(net_g, "module"):
+ ckpt = net_g.module.state_dict()
+ else:
+ ckpt = net_g.state_dict()
+ logger.info(
+ "saving final ckpt:%s"
+ % (
+ save_small_model(
+ ckpt, hps.sample_rate, hps.if_f0, hps.name, epoch, hps.version, hps
+ )
+ )
+ )
+ sleep(1)
+ os._exit(2333333)
+
+
+if __name__ == "__main__":
+ mp.set_start_method("spawn", force=True)
+ main()
diff --git a/infer/modules/uvr5/mdxnet.py b/infer/modules/uvr5/mdxnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..02e95a7544a17917557e7cd8c7e47b604161aefe
--- /dev/null
+++ b/infer/modules/uvr5/mdxnet.py
@@ -0,0 +1,243 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+from tqdm import tqdm
+import av
+
+from infer.lib.audio import downsample_audio
+
+cpu = torch.device("cpu")
+
+
+class ConvTDFNetTrim:
+ def __init__(
+ self, device, model_name, target_name, L, dim_f, dim_t, n_fft, hop=1024
+ ):
+ super(ConvTDFNetTrim, self).__init__()
+
+ self.dim_f = dim_f
+ self.dim_t = 2**dim_t
+ self.n_fft = n_fft
+ self.hop = hop
+ self.n_bins = self.n_fft // 2 + 1
+ self.chunk_size = hop * (self.dim_t - 1)
+ self.window = torch.hann_window(window_length=self.n_fft, periodic=True).to(
+ device
+ )
+ self.target_name = target_name
+ self.blender = "blender" in model_name
+
+ self.dim_c = 4
+ out_c = self.dim_c * 4 if target_name == "*" else self.dim_c
+ self.freq_pad = torch.zeros(
+ [1, out_c, self.n_bins - self.dim_f, self.dim_t]
+ ).to(device)
+
+ self.n = L // 2
+
+ def stft(self, x):
+ x = x.reshape([-1, self.chunk_size])
+ x = torch.stft(
+ x,
+ n_fft=self.n_fft,
+ hop_length=self.hop,
+ window=self.window,
+ center=True,
+ return_complex=True,
+ )
+ x = torch.view_as_real(x)
+ x = x.permute([0, 3, 1, 2])
+ x = x.reshape([-1, 2, 2, self.n_bins, self.dim_t]).reshape(
+ [-1, self.dim_c, self.n_bins, self.dim_t]
+ )
+ return x[:, :, : self.dim_f]
+
+ def istft(self, x, freq_pad=None):
+ freq_pad = (
+ self.freq_pad.repeat([x.shape[0], 1, 1, 1])
+ if freq_pad is None
+ else freq_pad
+ )
+ x = torch.cat([x, freq_pad], -2)
+ c = 4 * 2 if self.target_name == "*" else 2
+ x = x.reshape([-1, c, 2, self.n_bins, self.dim_t]).reshape(
+ [-1, 2, self.n_bins, self.dim_t]
+ )
+ x = x.permute([0, 2, 3, 1])
+ x = x.contiguous()
+ x = torch.view_as_complex(x)
+ x = torch.istft(
+ x, n_fft=self.n_fft, hop_length=self.hop, window=self.window, center=True
+ )
+ return x.reshape([-1, c, self.chunk_size])
+
+
+def get_models(device, dim_f, dim_t, n_fft):
+ return ConvTDFNetTrim(
+ device=device,
+ model_name="Conv-TDF",
+ target_name="vocals",
+ L=11,
+ dim_f=dim_f,
+ dim_t=dim_t,
+ n_fft=n_fft,
+ )
+
+
+class Predictor:
+ def __init__(self, args):
+ import onnxruntime as ort
+
+ logger.info(ort.get_available_providers())
+ self.args = args
+ self.model_ = get_models(
+ device=cpu, dim_f=args.dim_f, dim_t=args.dim_t, n_fft=args.n_fft
+ )
+ self.model = ort.InferenceSession(
+ os.path.join(args.onnx, self.model_.target_name + ".onnx"),
+ providers=[
+ "CUDAExecutionProvider",
+ "DmlExecutionProvider",
+ "CPUExecutionProvider",
+ ],
+ )
+ logger.info("ONNX load done")
+
+ def demix(self, mix):
+ samples = mix.shape[-1]
+ margin = self.args.margin
+ chunk_size = self.args.chunks * 44100
+ assert not margin == 0, "margin cannot be zero!"
+ if margin > chunk_size:
+ margin = chunk_size
+
+ segmented_mix = {}
+
+ if self.args.chunks == 0 or samples < chunk_size:
+ chunk_size = samples
+
+ counter = -1
+ for skip in range(0, samples, chunk_size):
+ counter += 1
+
+ s_margin = 0 if counter == 0 else margin
+ end = min(skip + chunk_size + margin, samples)
+
+ start = skip - s_margin
+
+ segmented_mix[skip] = mix[:, start:end].copy()
+ if end == samples:
+ break
+
+ sources = self.demix_base(segmented_mix, margin_size=margin)
+ """
+ mix:(2,big_sample)
+ segmented_mix:offset->(2,small_sample)
+ sources:(1,2,big_sample)
+ """
+ return sources
+
+ def demix_base(self, mixes, margin_size):
+ chunked_sources = []
+ progress_bar = tqdm(total=len(mixes))
+ progress_bar.set_description("Processing")
+ for mix in mixes:
+ cmix = mixes[mix]
+ sources = []
+ n_sample = cmix.shape[1]
+ model = self.model_
+ trim = model.n_fft // 2
+ gen_size = model.chunk_size - 2 * trim
+ pad = gen_size - n_sample % gen_size
+ mix_p = np.concatenate(
+ (np.zeros((2, trim)), cmix, np.zeros((2, pad)), np.zeros((2, trim))), 1
+ )
+ mix_waves = []
+ i = 0
+ while i < n_sample + pad:
+ waves = np.array(mix_p[:, i : i + model.chunk_size])
+ mix_waves.append(waves)
+ i += gen_size
+ mix_waves = torch.tensor(mix_waves, dtype=torch.float32).to(cpu)
+ with torch.no_grad():
+ _ort = self.model
+ spek = model.stft(mix_waves)
+ if self.args.denoise:
+ spec_pred = (
+ -_ort.run(None, {"input": -spek.cpu().numpy()})[0] * 0.5
+ + _ort.run(None, {"input": spek.cpu().numpy()})[0] * 0.5
+ )
+ tar_waves = model.istft(torch.tensor(spec_pred))
+ else:
+ tar_waves = model.istft(
+ torch.tensor(_ort.run(None, {"input": spek.cpu().numpy()})[0])
+ )
+ tar_signal = (
+ tar_waves[:, :, trim:-trim]
+ .transpose(0, 1)
+ .reshape(2, -1)
+ .numpy()[:, :-pad]
+ )
+
+ start = 0 if mix == 0 else margin_size
+ end = None if mix == list(mixes.keys())[::-1][0] else -margin_size
+ if margin_size == 0:
+ end = None
+ sources.append(tar_signal[:, start:end])
+
+ progress_bar.update(1)
+
+ chunked_sources.append(sources)
+ _sources = np.concatenate(chunked_sources, axis=-1)
+ # del self.model
+ progress_bar.close()
+ return _sources
+
+ def prediction(self, m, vocal_root, others_root, format):
+ os.makedirs(vocal_root, exist_ok=True)
+ os.makedirs(others_root, exist_ok=True)
+ basename = os.path.basename(m)
+ mix, rate = librosa.load(m, mono=False, sr=44100)
+ if mix.ndim == 1:
+ mix = np.asfortranarray([mix, mix])
+ mix = mix.T
+ sources = self.demix(mix.T)
+ opt = sources[0].T
+ if format in ["wav", "flac"]:
+ sf.write(
+ "%s/%s_main_vocal.%s" % (vocal_root, basename, format), mix - opt, rate
+ )
+ sf.write("%s/%s_others.%s" % (others_root, basename, format), opt, rate)
+ else:
+ path_vocal = "%s/%s_main_vocal.wav" % (vocal_root, basename)
+ path_other = "%s/%s_others.wav" % (others_root, basename)
+ sf.write(path_vocal, mix - opt, rate)
+ sf.write(path_other, opt, rate)
+ opt_path_vocal = path_vocal[:-4] + ".%s" % format
+ opt_path_other = path_other[:-4] + ".%s" % format
+ downsample_audio(path_vocal, opt_path_vocal, format)
+ downsample_audio(path_other, opt_path_other, format)
+
+
+class MDXNetDereverb:
+ def __init__(self, chunks, device):
+ self.onnx = "assets/uvr5_weights/onnx_dereverb_By_FoxJoy"
+ self.shifts = 10 # 'Predict with randomised equivariant stabilisation'
+ self.mixing = "min_mag" # ['default','min_mag','max_mag']
+ self.chunks = chunks
+ self.margin = 44100
+ self.dim_t = 9
+ self.dim_f = 3072
+ self.n_fft = 6144
+ self.denoise = True
+ self.pred = Predictor(self)
+ self.device = device
+
+ def _path_audio_(self, input, vocal_root, others_root, format, is_hp3=False):
+ self.pred.prediction(input, vocal_root, others_root, format)
diff --git a/infer/modules/uvr5/modules.py b/infer/modules/uvr5/modules.py
new file mode 100644
index 0000000000000000000000000000000000000000..dd070fe057afcad664a4a56e44258e0c6d5ad563
--- /dev/null
+++ b/infer/modules/uvr5/modules.py
@@ -0,0 +1,107 @@
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from infer.lib.audio import resample_audio, get_audio_properties
+import torch
+
+from configs import Config
+from infer.modules.uvr5.mdxnet import MDXNetDereverb
+from infer.modules.uvr5.vr import AudioPre, AudioPreDeEcho
+
+config = Config()
+
+
+def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
+ infos = []
+ try:
+ inp_root = inp_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+ save_root_vocal = (
+ save_root_vocal.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+ )
+ save_root_ins = (
+ save_root_ins.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+ )
+ if model_name == "onnx_dereverb_By_FoxJoy":
+ pre_fun = MDXNetDereverb(15, config.device)
+ else:
+ func = AudioPre if "DeEcho" not in model_name else AudioPreDeEcho
+ pre_fun = func(
+ agg=int(agg),
+ model_path=os.path.join(
+ os.getenv("weight_uvr5_root"), model_name + ".pth"
+ ),
+ device=config.device,
+ is_half=config.is_half,
+ )
+ is_hp3 = "HP3" in model_name
+ if inp_root != "":
+ paths = [os.path.join(inp_root, name) for name in os.listdir(inp_root)]
+ else:
+ paths = [path.name for path in paths]
+ for path in paths:
+ inp_path = os.path.join(inp_root, path)
+ need_reformat = 1
+ done = 0
+ try:
+ channels, rate = get_audio_properties(inp_path)
+
+ # Check the audio stream's properties
+ if channels == 2 and rate == 44100:
+ pre_fun._path_audio_(
+ inp_path, save_root_ins, save_root_vocal, format0, is_hp3=is_hp3
+ )
+ need_reformat = 0
+ done = 1
+ except Exception as e:
+ need_reformat = 1
+ print(f"Exception {e} occured. Will reformat")
+ if need_reformat == 1:
+ tmp_path = "%s/%s.reformatted.wav" % (
+ os.path.join(os.environ["TEMP"]),
+ os.path.basename(inp_path),
+ )
+ resample_audio(inp_path, tmp_path, "pcm_s16le", "s16", 44100, "stereo")
+ inp_path = tmp_path
+ try:
+ if done == 0:
+ pre_fun._path_audio_(
+ inp_path, save_root_ins, save_root_vocal, format0
+ )
+ infos.append("%s->Success" % (os.path.basename(inp_path)))
+ yield "\n".join(infos)
+ except:
+ try:
+ if done == 0:
+ pre_fun._path_audio_(
+ inp_path, save_root_ins, save_root_vocal, format0
+ )
+ infos.append("%s->Success" % (os.path.basename(inp_path)))
+ yield "\n".join(infos)
+ except:
+ infos.append(
+ "%s->%s" % (os.path.basename(inp_path), traceback.format_exc())
+ )
+ yield "\n".join(infos)
+ except:
+ infos.append(traceback.format_exc())
+ yield "\n".join(infos)
+ finally:
+ try:
+ if model_name == "onnx_dereverb_By_FoxJoy":
+ del pre_fun.pred.model
+ del pre_fun.pred.model_
+ else:
+ del pre_fun.model
+ del pre_fun
+ except:
+ traceback.print_exc()
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ logger.info("Executed torch.cuda.empty_cache()")
+ elif torch.backends.mps.is_available():
+ torch.mps.empty_cache()
+ logger.info("Executed torch.mps.empty_cache()")
+ yield "\n".join(infos)
diff --git a/infer/modules/uvr5/vr.py b/infer/modules/uvr5/vr.py
new file mode 100644
index 0000000000000000000000000000000000000000..85f192f00a9bd9ac2f1ecd4e7eec7417a443b175
--- /dev/null
+++ b/infer/modules/uvr5/vr.py
@@ -0,0 +1,348 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+from infer.lib.audio import downsample_audio
+import torch
+
+from infer.lib.uvr5_pack.lib_v5 import nets_123821KB as Nets
+from infer.lib.uvr5_pack.lib_v5 import spec_utils
+from infer.lib.uvr5_pack.lib_v5.model_param_init import ModelParameters
+from infer.lib.uvr5_pack.lib_v5.nets import CascadedNet
+from infer.lib.uvr5_pack.utils import inference
+
+
+class AudioPre:
+ def __init__(self, agg, model_path, device, is_half, tta=False):
+ self.model_path = model_path
+ self.device = device
+ self.data = {
+ # Processing Options
+ "postprocess": False,
+ "tta": tta,
+ # Constants
+ "window_size": 512,
+ "agg": agg,
+ "high_end_process": "mirroring",
+ }
+ mp = ModelParameters("infer/lib/uvr5_pack/lib_v5/modelparams/4band_v2.json")
+ model = Nets.CascadedASPPNet(mp.param["bins"] * 2)
+ cpk = torch.load(model_path, map_location="cpu")
+ model.load_state_dict(cpk)
+ model.eval()
+ if is_half:
+ model = model.half().to(device)
+ else:
+ model = model.to(device)
+
+ self.mp = mp
+ self.model = model
+
+ def _path_audio_(
+ self, music_file, ins_root=None, vocal_root=None, format="flac", is_hp3=False
+ ):
+ if ins_root is None and vocal_root is None:
+ return "No save root."
+ name = os.path.basename(music_file)
+ if ins_root is not None:
+ os.makedirs(ins_root, exist_ok=True)
+ if vocal_root is not None:
+ os.makedirs(vocal_root, exist_ok=True)
+ X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+ bands_n = len(self.mp.param["band"])
+ # print(bands_n)
+ for d in range(bands_n, 0, -1):
+ bp = self.mp.param["band"][d]
+ if d == bands_n: # high-end band
+ (
+ X_wave[d],
+ _,
+ ) = librosa.core.load( # 理论上librosa读取可能对某些音频有bug，应该上av读取，但是太麻烦了弃坑
+ music_file,
+ bp["sr"],
+ False,
+ dtype=np.float32,
+ res_type=bp["res_type"],
+ )
+ if X_wave[d].ndim == 1:
+ X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
+ else: # lower bands
+ X_wave[d] = librosa.core.resample(
+ X_wave[d + 1],
+ self.mp.param["band"][d + 1]["sr"],
+ bp["sr"],
+ res_type=bp["res_type"],
+ )
+ # Stft of wave source
+ X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
+ X_wave[d],
+ bp["hl"],
+ bp["n_fft"],
+ self.mp.param["mid_side"],
+ self.mp.param["mid_side_b2"],
+ self.mp.param["reverse"],
+ )
+ # pdb.set_trace()
+ if d == bands_n and self.data["high_end_process"] != "none":
+ input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
+ self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
+ )
+ input_high_end = X_spec_s[d][
+ :, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
+ ]
+
+ X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
+ aggresive_set = float(self.data["agg"] / 100)
+ aggressiveness = {
+ "value": aggresive_set,
+ "split_bin": self.mp.param["band"][1]["crop_stop"],
+ }
+ with torch.no_grad():
+ pred, X_mag, X_phase = inference(
+ X_spec_m, self.device, self.model, aggressiveness, self.data
+ )
+ # Postprocess
+ if self.data["postprocess"]:
+ pred_inv = np.clip(X_mag - pred, 0, np.inf)
+ pred = spec_utils.mask_silence(pred, pred_inv)
+ y_spec_m = pred * X_phase
+ v_spec_m = X_spec_m - y_spec_m
+
+ if ins_root is not None:
+ if self.data["high_end_process"].startswith("mirroring"):
+ input_high_end_ = spec_utils.mirroring(
+ self.data["high_end_process"], y_spec_m, input_high_end, self.mp
+ )
+ wav_instrument = spec_utils.cmb_spectrogram_to_wave(
+ y_spec_m, self.mp, input_high_end_h, input_high_end_
+ )
+ else:
+ wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
+ logger.info("%s instruments done" % name)
+ if is_hp3 == True:
+ head = "vocal_"
+ else:
+ head = "instrument_"
+ if format in ["wav", "flac"]:
+ sf.write(
+ os.path.join(
+ ins_root,
+ head + "{}_{}.{}".format(name, self.data["agg"], format),
+ ),
+ (np.array(wav_instrument) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ ) #
+ else:
+ path = os.path.join(
+ ins_root, head + "{}_{}.wav".format(name, self.data["agg"])
+ )
+ sf.write(
+ path,
+ (np.array(wav_instrument) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ if os.path.exists(path):
+ opt_format_path = path[:-4] + ".%s" % format
+ downsample_audio(path, opt_format_path, format)
+ if vocal_root is not None:
+ if is_hp3 == True:
+ head = "instrument_"
+ else:
+ head = "vocal_"
+ if self.data["high_end_process"].startswith("mirroring"):
+ input_high_end_ = spec_utils.mirroring(
+ self.data["high_end_process"], v_spec_m, input_high_end, self.mp
+ )
+ wav_vocals = spec_utils.cmb_spectrogram_to_wave(
+ v_spec_m, self.mp, input_high_end_h, input_high_end_
+ )
+ else:
+ wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
+ logger.info("%s vocals done" % name)
+ if format in ["wav", "flac"]:
+ sf.write(
+ os.path.join(
+ vocal_root,
+ head + "{}_{}.{}".format(name, self.data["agg"], format),
+ ),
+ (np.array(wav_vocals) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ else:
+ path = os.path.join(
+ vocal_root, head + "{}_{}.wav".format(name, self.data["agg"])
+ )
+ sf.write(
+ path,
+ (np.array(wav_vocals) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ opt_format_path = path[:-4] + ".%s" % format
+ downsample_audio(path, opt_format_path, format)
+
+
+class AudioPreDeEcho:
+ def __init__(self, agg, model_path, device, is_half, tta=False):
+ self.model_path = model_path
+ self.device = device
+ self.data = {
+ # Processing Options
+ "postprocess": False,
+ "tta": tta,
+ # Constants
+ "window_size": 512,
+ "agg": agg,
+ "high_end_process": "mirroring",
+ }
+ mp = ModelParameters("infer/lib/uvr5_pack/lib_v5/modelparams/4band_v3.json")
+ nout = 64 if "DeReverb" in model_path else 48
+ model = CascadedNet(mp.param["bins"] * 2, nout)
+ cpk = torch.load(model_path, map_location="cpu")
+ model.load_state_dict(cpk)
+ model.eval()
+ if is_half:
+ model = model.half().to(device)
+ else:
+ model = model.to(device)
+
+ self.mp = mp
+ self.model = model
+
+ def _path_audio_(
+ self, music_file, vocal_root=None, ins_root=None, format="flac", is_hp3=False
+ ): # 3个VR模型vocal和ins是反的
+ if ins_root is None and vocal_root is None:
+ return "No save root."
+ name = os.path.basename(music_file)
+ if ins_root is not None:
+ os.makedirs(ins_root, exist_ok=True)
+ if vocal_root is not None:
+ os.makedirs(vocal_root, exist_ok=True)
+ X_wave, y_wave, X_spec_s, y_spec_s = {}, {}, {}, {}
+ bands_n = len(self.mp.param["band"])
+ # print(bands_n)
+ for d in range(bands_n, 0, -1):
+ bp = self.mp.param["band"][d]
+ if d == bands_n: # high-end band
+ (
+ X_wave[d],
+ _,
+ ) = librosa.core.load( # 理论上librosa读取可能对某些音频有bug，应该上av读取，但是太麻烦了弃坑
+ music_file,
+ bp["sr"],
+ False,
+ dtype=np.float32,
+ res_type=bp["res_type"],
+ )
+ if X_wave[d].ndim == 1:
+ X_wave[d] = np.asfortranarray([X_wave[d], X_wave[d]])
+ else: # lower bands
+ X_wave[d] = librosa.core.resample(
+ X_wave[d + 1],
+ self.mp.param["band"][d + 1]["sr"],
+ bp["sr"],
+ res_type=bp["res_type"],
+ )
+ # Stft of wave source
+ X_spec_s[d] = spec_utils.wave_to_spectrogram_mt(
+ X_wave[d],
+ bp["hl"],
+ bp["n_fft"],
+ self.mp.param["mid_side"],
+ self.mp.param["mid_side_b2"],
+ self.mp.param["reverse"],
+ )
+ # pdb.set_trace()
+ if d == bands_n and self.data["high_end_process"] != "none":
+ input_high_end_h = (bp["n_fft"] // 2 - bp["crop_stop"]) + (
+ self.mp.param["pre_filter_stop"] - self.mp.param["pre_filter_start"]
+ )
+ input_high_end = X_spec_s[d][
+ :, bp["n_fft"] // 2 - input_high_end_h : bp["n_fft"] // 2, :
+ ]
+
+ X_spec_m = spec_utils.combine_spectrograms(X_spec_s, self.mp)
+ aggresive_set = float(self.data["agg"] / 100)
+ aggressiveness = {
+ "value": aggresive_set,
+ "split_bin": self.mp.param["band"][1]["crop_stop"],
+ }
+ with torch.no_grad():
+ pred, X_mag, X_phase = inference(
+ X_spec_m, self.device, self.model, aggressiveness, self.data
+ )
+ # Postprocess
+ if self.data["postprocess"]:
+ pred_inv = np.clip(X_mag - pred, 0, np.inf)
+ pred = spec_utils.mask_silence(pred, pred_inv)
+ y_spec_m = pred * X_phase
+ v_spec_m = X_spec_m - y_spec_m
+
+ if ins_root is not None:
+ if self.data["high_end_process"].startswith("mirroring"):
+ input_high_end_ = spec_utils.mirroring(
+ self.data["high_end_process"], y_spec_m, input_high_end, self.mp
+ )
+ wav_instrument = spec_utils.cmb_spectrogram_to_wave(
+ y_spec_m, self.mp, input_high_end_h, input_high_end_
+ )
+ else:
+ wav_instrument = spec_utils.cmb_spectrogram_to_wave(y_spec_m, self.mp)
+ logger.info("%s instruments done" % name)
+ if format in ["wav", "flac"]:
+ sf.write(
+ os.path.join(
+ ins_root,
+ "vocal_{}_{}.{}".format(name, self.data["agg"], format),
+ ),
+ (np.array(wav_instrument) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ ) #
+ else:
+ path = os.path.join(
+ ins_root, "vocal_{}_{}.wav".format(name, self.data["agg"])
+ )
+ sf.write(
+ path,
+ (np.array(wav_instrument) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ if os.path.exists(path):
+ opt_format_path = path[:-4] + ".%s" % format
+ downsample_audio(path, opt_format_path, format)
+ if vocal_root is not None:
+ if self.data["high_end_process"].startswith("mirroring"):
+ input_high_end_ = spec_utils.mirroring(
+ self.data["high_end_process"], v_spec_m, input_high_end, self.mp
+ )
+ wav_vocals = spec_utils.cmb_spectrogram_to_wave(
+ v_spec_m, self.mp, input_high_end_h, input_high_end_
+ )
+ else:
+ wav_vocals = spec_utils.cmb_spectrogram_to_wave(v_spec_m, self.mp)
+ logger.info("%s vocals done" % name)
+ if format in ["wav", "flac"]:
+ sf.write(
+ os.path.join(
+ vocal_root,
+ "instrument_{}_{}.{}".format(name, self.data["agg"], format),
+ ),
+ (np.array(wav_vocals) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ else:
+ path = os.path.join(
+ vocal_root, "instrument_{}_{}.wav".format(name, self.data["agg"])
+ )
+ sf.write(
+ path,
+ (np.array(wav_vocals) * 32768).astype("int16"),
+ self.mp.param["sr"],
+ )
+ if os.path.exists(path):
+ opt_format_path = path[:-4] + ".%s" % format
+ downsample_audio(path, opt_format_path, format)
diff --git a/infer/modules/vc/__init__.py b/infer/modules/vc/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..06b6084ebd7c36e4f0e6f02f7f8f96c9212055b7
--- /dev/null
+++ b/infer/modules/vc/__init__.py
@@ -0,0 +1,5 @@
+from .pipeline import Pipeline
+from .modules import VC
+from .utils import get_index_path_from_model, load_hubert
+from .info import show_info
+from .hash import model_hash_ckpt, hash_id, hash_similarity
diff --git a/infer/modules/vc/hash.py b/infer/modules/vc/hash.py
new file mode 100644
index 0000000000000000000000000000000000000000..d8d7e7c0f93ba20b13509a6e2e524cb5efba3efa
--- /dev/null
+++ b/infer/modules/vc/hash.py
@@ -0,0 +1,202 @@
+import numpy as np
+import torch
+import hashlib
+import pathlib
+from scipy.fft import fft
+from pybase16384 import encode_to_string, decode_from_string
+
+from configs import CPUConfig, singleton_variable
+from rvc.synthesizer import get_synthesizer
+
+from .pipeline import Pipeline
+from .utils import load_hubert
+
+
+class TorchSeedContext:
+ def __init__(self, seed):
+ self.seed = seed
+ self.state = None
+
+ def __enter__(self):
+ self.state = torch.random.get_rng_state()
+ torch.manual_seed(self.seed)
+
+ def __exit__(self, type, value, traceback):
+ torch.random.set_rng_state(self.state)
+
+
+half_hash_len = 512
+expand_factor = 65536 * 8
+
+
+@singleton_variable
+def original_audio_storage():
+ return np.load(pathlib.Path(__file__).parent / "lgdsng.npz")
+
+
+@singleton_variable
+def original_audio():
+ return original_audio_storage()["a"]
+
+
+@singleton_variable
+def original_audio_time_minus():
+ return original_audio_storage()["t"]
+
+
+@singleton_variable
+def original_audio_freq_minus():
+ return original_audio_storage()["f"]
+
+
+@singleton_variable
+def original_rmvpe_f0():
+ x = original_audio_storage()
+ return x["pitch"], x["pitchf"]
+
+
+def _cut_u16(n):
+ if n > 16384:
+ n = 16384 + 16384 * (1 - np.exp((16384 - n) / expand_factor))
+ elif n < -16384:
+ n = -16384 - 16384 * (1 - np.exp((n + 16384) / expand_factor))
+ return n
+
+
+# wave_hash will change time_field, use carefully
+def wave_hash(time_field):
+ np.divide(time_field, np.abs(time_field).max(), time_field)
+ if len(time_field) != 48000:
+ raise Exception("time not hashable")
+ freq_field = fft(time_field)
+ if len(freq_field) != 48000:
+ raise Exception("freq not hashable")
+ np.add(time_field, original_audio_time_minus(), out=time_field)
+ np.add(freq_field, original_audio_freq_minus(), out=freq_field)
+ hash = np.zeros(half_hash_len // 2 * 2, dtype=">i2")
+ d = 375 * 512 // half_hash_len
+ for i in range(half_hash_len // 4):
+ a = i * 2
+ b = a + 1
+ x = a + half_hash_len // 2
+ y = x + 1
+ s = np.average(freq_field[i * d : (i + 1) * d])
+ hash[a] = np.int16(_cut_u16(round(32768 * np.real(s))))
+ hash[b] = np.int16(_cut_u16(round(32768 * np.imag(s))))
+ hash[x] = np.int16(
+ _cut_u16(round(32768 * np.sum(time_field[i * d : i * d + d // 2])))
+ )
+ hash[y] = np.int16(
+ _cut_u16(round(32768 * np.sum(time_field[i * d + d // 2 : (i + 1) * d])))
+ )
+ return encode_to_string(hash.tobytes())
+
+
+def model_hash(config, tgt_sr, net_g, if_f0, version):
+ pipeline = Pipeline(tgt_sr, config)
+ audio = original_audio()
+ hbt = load_hubert(config.device, config.is_half)
+ audio_opt = pipeline.pipeline(
+ hbt,
+ net_g,
+ 0,
+ audio,
+ [0, 0, 0],
+ 6,
+ original_rmvpe_f0(),
+ "",
+ 0,
+ 2 if if_f0 else 0,
+ 3,
+ tgt_sr,
+ 16000,
+ 0.25,
+ version,
+ 0.33,
+ )
+ del hbt
+ opt_len = len(audio_opt)
+ diff = 48000 - opt_len
+ if diff > 0:
+ audio_opt = np.pad(audio_opt, (diff, 0))
+ elif diff < 0:
+ n = diff // 2
+ n = -n
+ audio_opt = audio_opt[n:-n]
+ h = wave_hash(audio_opt)
+ del pipeline, audio_opt
+ return h
+
+
+def model_hash_ckpt(cpt):
+ config = CPUConfig()
+
+ with TorchSeedContext(114514):
+ net_g, cpt = get_synthesizer(cpt, config.device)
+ tgt_sr = cpt["config"][-1]
+ if_f0 = cpt.get("f0", 1)
+ version = cpt.get("version", "v1")
+
+ if config.is_half:
+ net_g = net_g.half()
+ else:
+ net_g = net_g.float()
+
+ h = model_hash(config, tgt_sr, net_g, if_f0, version)
+
+ del net_g
+
+ return h
+
+
+def model_hash_from(path):
+ cpt = torch.load(path, map_location="cpu")
+ h = model_hash_ckpt(cpt)
+ del cpt
+ return h
+
+
+def _extend_difference(n, a, b):
+ if n < a:
+ n = a
+ elif n > b:
+ n = b
+ n -= a
+ n /= b - a
+ return n
+
+
+def hash_similarity(h1: str, h2: str) -> float:
+ try:
+ h1b, h2b = decode_from_string(h1), decode_from_string(h2)
+ if len(h1b) != half_hash_len * 2 or len(h2b) != half_hash_len * 2:
+ raise Exception("invalid hash length")
+ h1n, h2n = np.frombuffer(h1b, dtype=">i2"), np.frombuffer(h2b, dtype=">i2")
+ d = 0
+ for i in range(half_hash_len // 4):
+ a = i * 2
+ b = a + 1
+ ax = complex(h1n[a], h1n[b])
+ bx = complex(h2n[a], h2n[b])
+ if abs(ax) == 0 or abs(bx) == 0:
+ continue
+ d += np.abs(ax - bx)
+ frac = np.linalg.norm(h1n) * np.linalg.norm(h2n)
+ cosine = (
+ np.dot(h1n.astype(np.float32), h2n.astype(np.float32)) / frac
+ if frac != 0
+ else 1.0
+ )
+ distance = _extend_difference(np.exp(-d / expand_factor), 0.5, 1.0)
+ return round((abs(cosine) + distance) / 2, 6)
+ except Exception as e:
+ return str(e)
+
+
+def hash_id(h: str) -> str:
+ d = decode_from_string(h)
+ if len(d) != half_hash_len * 2:
+ return "invalid hash length"
+ return encode_to_string(
+ np.frombuffer(d, dtype=np.uint64).sum(keepdims=True).tobytes()
+ )[:-2] + encode_to_string(hashlib.md5(d).digest()[:7])
diff --git a/infer/modules/vc/info.py b/infer/modules/vc/info.py
new file mode 100644
index 0000000000000000000000000000000000000000..14a734634384865629ec1863d9b68a1d02228884
--- /dev/null
+++ b/infer/modules/vc/info.py
@@ -0,0 +1,84 @@
+import traceback
+from i18n.i18n import I18nAuto
+from datetime import datetime
+import torch
+
+from .hash import model_hash_ckpt, hash_id, hash_similarity
+
+i18n = I18nAuto()
+
+
+def show_model_info(cpt, show_long_id=False):
+ try:
+ h = model_hash_ckpt(cpt)
+ id = hash_id(h)
+ idread = cpt.get("id", "None")
+ hread = cpt.get("hash", "None")
+ if id != idread:
+ id += (
+ "("
+ + i18n("Actually calculated")
+ + "), "
+ + idread
+ + "("
+ + i18n("Read from model")
+ + ")"
+ )
+ sim = hash_similarity(h, hread)
+ if not isinstance(sim, str):
+ sim = "%.2f%%" % (sim * 100)
+ if not show_long_id:
+ h = i18n("Hidden")
+ if h != hread:
+ h = i18n("Similarity") + " " + sim + " -> " + h
+ elif h != hread:
+ h = (
+ i18n("Similarity")
+ + " "
+ + sim
+ + " -> "
+ + h
+ + "("
+ + i18n("Actually calculated")
+ + "), "
+ + hread
+ + "("
+ + i18n("Read from model")
+ + ")"
+ )
+ txt = f"""{i18n("Model name")}: %s
+{i18n("Sealing date")}: %s
+{i18n("Model Author")}: %s
+{i18n("Information")}: %s
+{i18n("Sampling rate")}: %s
+{i18n("Pitch guidance (f0)")}: %s
+{i18n("Version")}: %s
+{i18n("ID(short)")}: %s
+{i18n("ID(long)")}: %s""" % (
+ cpt.get("name", i18n("Unknown")),
+ datetime.fromtimestamp(float(cpt.get("timestamp", 0))),
+ cpt.get("author", i18n("Unknown")),
+ cpt.get("info", i18n("None")),
+ cpt.get("sr", i18n("Unknown")),
+ i18n("Exist") if cpt.get("f0", 0) == 1 else i18n("Not exist"),
+ cpt.get("version", i18n("None")),
+ id,
+ h,
+ )
+ except:
+ txt = traceback.format_exc()
+
+ return txt
+
+
+def show_info(path):
+ try:
+ if hasattr(path, "name"):
+ path = path.name
+ a = torch.load(path, map_location="cpu")
+ txt = show_model_info(a, show_long_id=True)
+ del a
+ except:
+ txt = traceback.format_exc()
+
+ return txt
diff --git a/infer/modules/vc/lgdsng.npz b/infer/modules/vc/lgdsng.npz
new file mode 100644
index 0000000000000000000000000000000000000000..eec46f1df99730f6d6662eccbb854c9a4015a270
--- /dev/null
+++ b/infer/modules/vc/lgdsng.npz
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6c2cea46a96b75c3e3154486672c99b335549403f9f61c77ae5bb22854950864
+size 708982
diff --git a/infer/modules/vc/modules.py b/infer/modules/vc/modules.py
new file mode 100644
index 0000000000000000000000000000000000000000..49f21dc8d1a4f9946e1966dc573f69cf7673aa0d
--- /dev/null
+++ b/infer/modules/vc/modules.py
@@ -0,0 +1,279 @@
+import traceback
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+import numpy as np
+import soundfile as sf
+import torch
+from io import BytesIO
+
+from infer.lib.audio import load_audio, wav2
+from rvc.synthesizer import get_synthesizer, load_synthesizer
+from .info import show_model_info
+from .pipeline import Pipeline
+from .utils import get_index_path_from_model, load_hubert
+
+
+class VC:
+ def __init__(self, config):
+ self.n_spk = None
+ self.tgt_sr = None
+ self.net_g = None
+ self.pipeline = None
+ self.cpt = None
+ self.version = None
+ self.if_f0 = None
+ self.version = None
+ self.hubert_model = None
+
+ self.config = config
+
+ def get_vc(self, sid, *to_return_protect):
+ logger.info("Get sid: " + sid)
+
+ to_return_protect0 = {
+ "visible": self.if_f0 != 0,
+ "value": (
+ to_return_protect[0] if self.if_f0 != 0 and to_return_protect else 0.5
+ ),
+ "__type__": "update",
+ }
+ to_return_protect1 = {
+ "visible": self.if_f0 != 0,
+ "value": (
+ to_return_protect[1] if self.if_f0 != 0 and to_return_protect else 0.33
+ ),
+ "__type__": "update",
+ }
+
+ if sid == "" or sid == []:
+ if (
+ self.hubert_model is not None
+ ): # 考虑到轮询, 需要加个判断看是否 sid 是由有模型切换到无模型的
+ logger.info("Clean model cache")
+ del (self.net_g, self.n_spk, self.hubert_model, self.tgt_sr) # ,cpt
+ self.hubert_model = self.net_g = self.n_spk = self.hubert_model = (
+ self.tgt_sr
+ ) = None
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ elif torch.backends.mps.is_available():
+ torch.mps.empty_cache()
+ ###楼下不这么折腾清理不干净
+ self.net_g, self.cpt = get_synthesizer(self.cpt, self.config.device)
+ self.if_f0 = self.cpt.get("f0", 1)
+ self.version = self.cpt.get("version", "v1")
+ del self.net_g, self.cpt
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ elif torch.backends.mps.is_available():
+ torch.mps.empty_cache()
+ return (
+ (
+ {"visible": False, "__type__": "update"},
+ to_return_protect0,
+ to_return_protect1,
+ {"value": to_return_protect[2], "__type__": "update"},
+ {"value": to_return_protect[3], "__type__": "update"},
+ {"value": "", "__type__": "update"},
+ )
+ if to_return_protect
+ else {"visible": True, "maximum": 0, "__type__": "update"}
+ )
+
+ person = f'{os.getenv("weight_root")}/{sid}'
+ logger.info(f"Loading: {person}")
+
+ self.net_g, self.cpt = load_synthesizer(person, self.config.device)
+ self.tgt_sr = self.cpt["config"][-1]
+ self.cpt["config"][-3] = self.cpt["weight"]["emb_g.weight"].shape[0] # n_spk
+ self.if_f0 = self.cpt.get("f0", 1)
+ self.version = self.cpt.get("version", "v1")
+
+ if self.config.is_half:
+ self.net_g = self.net_g.half()
+ else:
+ self.net_g = self.net_g.float()
+ self.pipeline = Pipeline(self.tgt_sr, self.config)
+
+ n_spk = self.cpt["config"][-3]
+ index = {"value": get_index_path_from_model(sid), "__type__": "update"}
+ logger.info("Select index: " + index["value"])
+
+ return (
+ (
+ {"visible": True, "maximum": n_spk, "__type__": "update"},
+ to_return_protect0,
+ to_return_protect1,
+ index,
+ index,
+ show_model_info(self.cpt),
+ )
+ if to_return_protect
+ else {"visible": True, "maximum": n_spk, "__type__": "update"}
+ )
+
+ def vc_single(
+ self,
+ sid,
+ input_audio_path,
+ f0_up_key,
+ f0_file,
+ f0_method,
+ file_index,
+ file_index2,
+ index_rate,
+ filter_radius,
+ resample_sr,
+ rms_mix_rate,
+ protect,
+ ):
+ if input_audio_path is None:
+ return "You need to upload an audio", None
+ elif hasattr(input_audio_path, "name"):
+ input_audio_path = str(input_audio_path.name)
+ f0_up_key = int(f0_up_key)
+ try:
+ audio = load_audio(input_audio_path, 16000)
+ audio_max = np.abs(audio).max() / 0.95
+ if audio_max > 1:
+ np.divide(audio, audio_max, audio)
+ times = [0, 0, 0]
+
+ if self.hubert_model is None:
+ self.hubert_model = load_hubert(self.config.device, self.config.is_half)
+
+ if file_index:
+ if hasattr(file_index, "name"):
+ file_index = str(file_index.name)
+ file_index = (
+ file_index.strip(" ")
+ .strip('"')
+ .strip("\n")
+ .strip('"')
+ .strip(" ")
+ .replace("trained", "added")
+ )
+ elif file_index2:
+ file_index = file_index2
+ else:
+ file_index = "" # 防止小白写错，自动帮他替换掉
+
+ audio_opt = self.pipeline.pipeline(
+ self.hubert_model,
+ self.net_g,
+ sid,
+ audio,
+ times,
+ f0_up_key,
+ f0_method,
+ file_index,
+ index_rate,
+ self.if_f0,
+ filter_radius,
+ self.tgt_sr,
+ resample_sr,
+ rms_mix_rate,
+ self.version,
+ protect,
+ f0_file,
+ ).astype(np.int16)
+ if self.tgt_sr != resample_sr >= 16000:
+ tgt_sr = resample_sr
+ else:
+ tgt_sr = self.tgt_sr
+ index_info = (
+ "Index: %s." % file_index
+ if os.path.exists(file_index)
+ else "Index not used."
+ )
+ return (
+ "Success.\n%s\nTime: npy: %.2fs, f0: %.2fs, infer: %.2fs."
+ % (index_info, *times),
+ (tgt_sr, audio_opt),
+ )
+ except Exception as e:
+ info = traceback.format_exc()
+ logger.warning(info)
+ return str(e), None
+
+ def vc_multi(
+ self,
+ sid,
+ dir_path,
+ opt_root,
+ paths,
+ f0_up_key,
+ f0_method,
+ file_index,
+ file_index2,
+ index_rate,
+ filter_radius,
+ resample_sr,
+ rms_mix_rate,
+ protect,
+ format1,
+ ):
+ try:
+ dir_path = (
+ dir_path.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+ ) # 防止小白拷路径头尾带了空格和"和回车
+ opt_root = opt_root.strip(" ").strip('"').strip("\n").strip('"').strip(" ")
+ os.makedirs(opt_root, exist_ok=True)
+ try:
+ if dir_path != "":
+ paths = [
+ os.path.join(dir_path, name) for name in os.listdir(dir_path)
+ ]
+ else:
+ paths = [path.name for path in paths]
+ except:
+ traceback.print_exc()
+ paths = [path.name for path in paths]
+ infos = []
+ for path in paths:
+ info, opt = self.vc_single(
+ sid,
+ path,
+ f0_up_key,
+ None,
+ f0_method,
+ file_index,
+ file_index2,
+ # file_big_npy,
+ index_rate,
+ filter_radius,
+ resample_sr,
+ rms_mix_rate,
+ protect,
+ )
+ if "Success" in info:
+ try:
+ tgt_sr, audio_opt = opt
+ if format1 in ["wav", "flac"]:
+ sf.write(
+ "%s/%s.%s"
+ % (opt_root, os.path.basename(path), format1),
+ audio_opt,
+ tgt_sr,
+ )
+ else:
+ path = "%s/%s.%s" % (
+ opt_root,
+ os.path.basename(path),
+ format1,
+ )
+ with BytesIO() as wavf:
+ sf.write(wavf, audio_opt, tgt_sr, format="wav")
+ wavf.seek(0, 0)
+ with open(path, "wb") as outf:
+ wav2(wavf, outf, format1)
+ except:
+ info += traceback.format_exc()
+ infos.append("%s->%s" % (os.path.basename(path), info))
+ yield "\n".join(infos)
+ yield "\n".join(infos)
+ except:
+ yield traceback.format_exc()
diff --git a/infer/modules/vc/pipeline.py b/infer/modules/vc/pipeline.py
new file mode 100644
index 0000000000000000000000000000000000000000..4f4a401405b2549669853936a29bf9abf168c792
--- /dev/null
+++ b/infer/modules/vc/pipeline.py
@@ -0,0 +1,444 @@
+import os
+import sys
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from time import time
+
+import faiss
+import librosa
+import numpy as np
+import torch
+import torch.nn.functional as F
+from scipy import signal
+
+from rvc.f0 import PM, Harvest, RMVPE, CRePE, Dio, FCPE
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
+
+
+def change_rms(data1, sr1, data2, sr2, rate): # 1是输入音频，2是输出音频,rate是2的占比
+ # print(data1.max(),data2.max())
+ rms1 = librosa.feature.rms(
+ y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
+ ) # 每半秒一个点
+ rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
+ rms1 = torch.from_numpy(rms1)
+ rms1 = F.interpolate(
+ rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
+ ).squeeze()
+ rms2 = torch.from_numpy(rms2)
+ rms2 = F.interpolate(
+ rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
+ ).squeeze()
+ rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
+ data2 *= (
+ torch.pow(rms1, torch.tensor(1 - rate))
+ * torch.pow(rms2, torch.tensor(rate - 1))
+ ).numpy()
+ return data2
+
+
+class Pipeline(object):
+ def __init__(self, tgt_sr, config):
+ self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
+ config.x_pad,
+ config.x_query,
+ config.x_center,
+ config.x_max,
+ config.is_half,
+ )
+ self.sr = 16000 # hubert输入采样率
+ self.window = 160 # 每帧点数
+ self.t_pad = self.sr * self.x_pad # 每条前后pad时间
+ self.t_pad_tgt = tgt_sr * self.x_pad
+ self.t_pad2 = self.t_pad * 2
+ self.t_query = self.sr * self.x_query # 查询切点前后查询时间
+ self.t_center = self.sr * self.x_center # 查询切点位置
+ self.t_max = self.sr * self.x_max # 免查询时长阈值
+ self.device = config.device
+
+ def get_f0(
+ self,
+ x,
+ p_len,
+ f0_up_key,
+ f0_method,
+ filter_radius,
+ inp_f0=None,
+ ):
+ f0_min = 50
+ f0_max = 1100
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+ if f0_method == "pm":
+ if not hasattr(self, "pm"):
+ self.pm = PM(self.window, f0_min, f0_max, self.sr)
+ f0 = self.pm.compute_f0(x, p_len=p_len)
+ if f0_method == "dio":
+ if not hasattr(self, "dio"):
+ self.dio = Dio(self.window, f0_min, f0_max, self.sr)
+ f0 = self.dio.compute_f0(x, p_len=p_len)
+ elif f0_method == "harvest":
+ if not hasattr(self, "harvest"):
+ self.harvest = Harvest(self.window, f0_min, f0_max, self.sr)
+ f0 = self.harvest.compute_f0(x, p_len=p_len, filter_radius=filter_radius)
+ elif f0_method == "crepe":
+ if not hasattr(self, "crepe"):
+ self.crepe = CRePE(
+ self.window,
+ f0_min,
+ f0_max,
+ self.sr,
+ self.device,
+ )
+ f0 = self.crepe.compute_f0(x, p_len=p_len)
+ elif f0_method == "rmvpe":
+ if not hasattr(self, "rmvpe"):
+ logger.info(
+ "Loading rmvpe model %s" % "%s/rmvpe.pt" % os.environ["rmvpe_root"]
+ )
+ self.rmvpe = RMVPE(
+ "%s/rmvpe.pt" % os.environ["rmvpe_root"],
+ is_half=self.is_half,
+ device=self.device,
+ # use_jit=self.config.use_jit,
+ )
+ f0 = self.rmvpe.compute_f0(x, p_len=p_len, filter_radius=0.03)
+
+ if "privateuseone" in str(self.device): # clean ortruntime memory
+ del self.rmvpe.model
+ del self.rmvpe
+ logger.info("Cleaning ortruntime memory")
+
+ elif f0_method == "fcpe":
+ if not hasattr(self, "model_fcpe"):
+ logger.info("Loading fcpe model")
+ self.model_fcpe = FCPE(
+ self.window,
+ f0_min,
+ f0_max,
+ self.sr,
+ self.device,
+ )
+ f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
+
+ f0 *= pow(2, f0_up_key / 12)
+ # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+ tf0 = self.sr // self.window # 每秒f0点数
+ if inp_f0 is not None:
+ delta_t = np.round(
+ (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
+ ).astype("int16")
+ replace_f0 = np.interp(
+ list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
+ )
+ shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
+ f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
+ :shape
+ ]
+ # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
+ f0bak = f0.copy()
+ f0_mel = 1127 * np.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+ f0_mel_max - f0_mel_min
+ ) + 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > 255] = 255
+ f0_coarse = np.rint(f0_mel).astype(np.int32)
+ return f0_coarse, f0bak # 1-0
+
+ def vc(
+ self,
+ model,
+ net_g,
+ sid,
+ audio0,
+ pitch,
+ pitchf,
+ times,
+ index,
+ big_npy,
+ index_rate,
+ version,
+ protect,
+ ): # ,file_index,file_big_npy
+ feats = torch.from_numpy(audio0)
+ if self.is_half:
+ feats = feats.half()
+ else:
+ feats = feats.float()
+ if feats.dim() == 2: # double channels
+ feats = feats.mean(-1)
+ assert feats.dim() == 1, feats.dim()
+ feats = feats.view(1, -1)
+ padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)
+
+ inputs = {
+ "source": feats.to(self.device),
+ "padding_mask": padding_mask,
+ "output_layer": 9 if version == "v1" else 12,
+ }
+ t0 = time()
+ with torch.no_grad():
+ logits = model.extract_features(**inputs)
+ feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
+ if protect < 0.5 and pitch is not None and pitchf is not None:
+ feats0 = feats.clone()
+ if (
+ not isinstance(index, type(None))
+ and not isinstance(big_npy, type(None))
+ and index_rate != 0
+ ):
+ npy = feats[0].cpu().numpy()
+ if self.is_half:
+ npy = npy.astype("float32")
+
+ # _, I = index.search(npy, 1)
+ # npy = big_npy[I.squeeze()]
+
+ try:
+ score, ix = index.search(npy, k=8)
+ except:
+ raise Exception("index mistatch")
+ weight = np.square(1 / score)
+ weight /= weight.sum(axis=1, keepdims=True)
+ npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
+
+ if self.is_half:
+ npy = npy.astype("float16")
+ feats = (
+ torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
+ + (1 - index_rate) * feats
+ )
+
+ feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+ if protect < 0.5 and pitch is not None and pitchf is not None:
+ feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(
+ 0, 2, 1
+ )
+ t1 = time()
+ p_len = audio0.shape[0] // self.window
+ if feats.shape[1] < p_len:
+ p_len = feats.shape[1]
+ if pitch is not None and pitchf is not None:
+ pitch = pitch[:, :p_len]
+ pitchf = pitchf[:, :p_len]
+
+ if protect < 0.5 and pitch is not None and pitchf is not None:
+ pitchff = pitchf.clone()
+ pitchff[pitchf > 0] = 1
+ pitchff[pitchf < 1] = protect
+ pitchff = pitchff.unsqueeze(-1)
+ feats = feats * pitchff + feats0 * (1 - pitchff)
+ feats = feats.to(feats0.dtype)
+ p_len = torch.tensor([p_len], device=self.device).long()
+ with torch.no_grad():
+ audio1 = (
+ (
+ net_g.infer(
+ feats,
+ p_len,
+ sid,
+ pitch=pitch,
+ pitchf=pitchf,
+ )[0, 0]
+ )
+ .data.cpu()
+ .float()
+ .numpy()
+ )
+ del feats, p_len, padding_mask
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ elif torch.backends.mps.is_available():
+ torch.mps.empty_cache()
+ t2 = time()
+ times[0] += t1 - t0
+ times[2] += t2 - t1
+ return audio1
+
+ def pipeline(
+ self,
+ model,
+ net_g,
+ sid,
+ audio,
+ times,
+ f0_up_key,
+ f0_method,
+ file_index,
+ index_rate,
+ if_f0,
+ filter_radius,
+ tgt_sr,
+ resample_sr,
+ rms_mix_rate,
+ version,
+ protect,
+ f0_file=None,
+ ):
+ if (
+ file_index != ""
+ # and file_big_npy != ""
+ # and os.path.exists(file_big_npy) == True
+ and os.path.exists(file_index)
+ and index_rate != 0
+ ):
+ try:
+ index = faiss.read_index(file_index)
+ big_npy = index.reconstruct_n(0, index.ntotal)
+ except:
+ traceback.print_exc()
+ index = big_npy = None
+ else:
+ index = big_npy = None
+ audio = signal.filtfilt(bh, ah, audio)
+ audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
+ opt_ts = []
+ if audio_pad.shape[0] > self.t_max:
+ audio_sum = np.zeros_like(audio)
+ for i in range(self.window):
+ audio_sum += np.abs(audio_pad[i : i - self.window])
+ for t in range(self.t_center, audio.shape[0], self.t_center):
+ opt_ts.append(
+ t
+ - self.t_query
+ + np.where(
+ audio_sum[t - self.t_query : t + self.t_query]
+ == audio_sum[t - self.t_query : t + self.t_query].min()
+ )[0][0]
+ )
+ s = 0
+ audio_opt = []
+ t = None
+ t1 = time()
+ audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
+ p_len = audio_pad.shape[0] // self.window
+ inp_f0 = None
+ if hasattr(f0_file, "name"):
+ try:
+ with open(f0_file.name, "r") as f:
+ lines = f.read().strip("\n").split("\n")
+ inp_f0 = []
+ for line in lines:
+ inp_f0.append([float(i) for i in line.split(",")])
+ inp_f0 = np.array(inp_f0, dtype="float32")
+ except:
+ traceback.print_exc()
+ sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
+ pitch, pitchf = None, None
+ if if_f0:
+ if if_f0 == 1:
+ pitch, pitchf = self.get_f0(
+ audio_pad,
+ p_len,
+ f0_up_key,
+ f0_method,
+ filter_radius,
+ inp_f0,
+ )
+ elif if_f0 == 2:
+ pitch, pitchf = f0_method
+ pitch = pitch[:p_len]
+ pitchf = pitchf[:p_len]
+ if "mps" not in str(self.device) or "xpu" not in str(self.device):
+ pitchf = pitchf.astype(np.float32)
+ pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
+ pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
+ t2 = time()
+ times[1] += t2 - t1
+ for t in opt_ts:
+ t = t // self.window * self.window
+ if if_f0:
+ audio_opt.append(
+ self.vc(
+ model,
+ net_g,
+ sid,
+ audio_pad[s : t + self.t_pad2 + self.window],
+ pitch[:, s // self.window : (t + self.t_pad2) // self.window],
+ pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
+ times,
+ index,
+ big_npy,
+ index_rate,
+ version,
+ protect,
+ )[self.t_pad_tgt : -self.t_pad_tgt]
+ )
+ else:
+ audio_opt.append(
+ self.vc(
+ model,
+ net_g,
+ sid,
+ audio_pad[s : t + self.t_pad2 + self.window],
+ None,
+ None,
+ times,
+ index,
+ big_npy,
+ index_rate,
+ version,
+ protect,
+ )[self.t_pad_tgt : -self.t_pad_tgt]
+ )
+ s = t
+ if if_f0:
+ audio_opt.append(
+ self.vc(
+ model,
+ net_g,
+ sid,
+ audio_pad[t:],
+ pitch[:, t // self.window :] if t is not None else pitch,
+ pitchf[:, t // self.window :] if t is not None else pitchf,
+ times,
+ index,
+ big_npy,
+ index_rate,
+ version,
+ protect,
+ )[self.t_pad_tgt : -self.t_pad_tgt]
+ )
+ else:
+ audio_opt.append(
+ self.vc(
+ model,
+ net_g,
+ sid,
+ audio_pad[t:],
+ None,
+ None,
+ times,
+ index,
+ big_npy,
+ index_rate,
+ version,
+ protect,
+ )[self.t_pad_tgt : -self.t_pad_tgt]
+ )
+ audio_opt = np.concatenate(audio_opt)
+ if rms_mix_rate != 1:
+ audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
+ if tgt_sr != resample_sr >= 16000:
+ audio_opt = librosa.resample(
+ audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
+ )
+ audio_max = np.abs(audio_opt).max() / 0.99
+ max_int16 = 32768
+ if audio_max > 1:
+ max_int16 /= audio_max
+ np.multiply(audio_opt, max_int16, audio_opt)
+ del pitch, pitchf, sid
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ elif torch.backends.mps.is_available():
+ torch.mps.empty_cache()
+ return audio_opt
diff --git a/infer/modules/vc/utils.py b/infer/modules/vc/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0df05f623137056f887ae58efed0f53bb7bb9c9b
--- /dev/null
+++ b/infer/modules/vc/utils.py
@@ -0,0 +1,34 @@
+import os
+
+from fairseq import checkpoint_utils
+
+
+def get_index_path_from_model(sid):
+ return next(
+ (
+ f
+ for f in [
+ os.path.join(root, name)
+ for path in [os.getenv("outside_index_root"), os.getenv("index_root")]
+ for root, _, files in os.walk(path, topdown=False)
+ for name in files
+ if name.endswith(".index") and "trained" not in name
+ ]
+ if sid.split(".")[0] in f
+ ),
+ "",
+ )
+
+
+def load_hubert(device, is_half):
+ models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
+ ["assets/hubert/hubert_base.pt"],
+ suffix="",
+ )
+ hubert_model = models[0]
+ hubert_model = hubert_model.to(device)
+ if is_half:
+ hubert_model = hubert_model.half()
+ else:
+ hubert_model = hubert_model.float()
+ return hubert_model.eval()
diff --git a/requirements/amd.txt b/requirements/amd.txt
new file mode 100644
index 0000000000000000000000000000000000000000..ca5522b41a18211037c50cde9494c388db52d6fc
--- /dev/null
+++ b/requirements/amd.txt
@@ -0,0 +1,48 @@
+tensorflow-rocm
+joblib>=1.1.0
+numba==0.56.4
+numpy==1.23.5
+scipy
+librosa==0.9.1
+llvmlite==0.39.0
+fairseq==0.12.2
+faiss-cpu==1.7.3
+gradio
+Cython
+pydub>=0.25.1
+soundfile>=0.12.1
+tensorboardX
+Jinja2>=3.1.2
+json5
+Markdown
+matplotlib>=3.7.0
+matplotlib-inline>=0.1.3
+praat-parselmouth>=0.4.2
+Pillow>=9.1.1
+resampy>=0.4.2
+scikit-learn
+tensorboard
+tqdm>=4.63.1
+tornado>=6.1
+Werkzeug>=2.2.3
+uc-micro-py>=1.0.1
+sympy>=1.11.1
+tabulate>=0.8.10
+PyYAML>=6.0
+pyasn1>=0.4.8
+pyasn1-modules>=0.2.8
+fsspec>=2022.11.0
+absl-py>=1.2.0
+audioread
+uvicorn>=0.21.1
+colorama>=0.4.5
+pyworld==0.3.2
+httpx
+onnxruntime
+onnxruntime-gpu
+torchcrepe==0.0.20
+fastapi
+python-dotenv>=1.0.0
+av
+torchfcpe
+pybase16384
diff --git a/requirements/dml.txt b/requirements/dml.txt
new file mode 100644
index 0000000000000000000000000000000000000000..e261d0616d41a56239c00725209bea3983aaedeb
--- /dev/null
+++ b/requirements/dml.txt
@@ -0,0 +1,46 @@
+joblib>=1.1.0
+numba==0.56.4
+numpy==1.23.5
+scipy
+librosa==0.9.1
+llvmlite==0.39.0
+fairseq==0.12.2
+faiss-cpu==1.7.3
+gradio
+Cython
+pydub>=0.25.1
+soundfile>=0.12.1
+tensorboardX
+Jinja2>=3.1.2
+json5
+Markdown
+matplotlib>=3.7.0
+matplotlib-inline>=0.1.3
+praat-parselmouth>=0.4.2
+Pillow>=9.1.1
+resampy>=0.4.2
+scikit-learn
+tensorboard
+tqdm>=4.63.1
+tornado>=6.1
+Werkzeug>=2.2.3
+uc-micro-py>=1.0.1
+sympy>=1.11.1
+tabulate>=0.8.10
+PyYAML>=6.0
+pyasn1>=0.4.8
+pyasn1-modules>=0.2.8
+fsspec>=2022.11.0
+absl-py>=1.2.0
+audioread
+uvicorn>=0.21.1
+colorama>=0.4.5
+pyworld==0.3.2
+httpx
+onnxruntime-directml
+torchcrepe==0.0.20
+fastapi
+python-dotenv>=1.0.0
+av
+torchfcpe
+pybase16384
diff --git a/requirements/gui-dml.txt b/requirements/gui-dml.txt
new file mode 100644
index 0000000000000000000000000000000000000000..6f7e11eae099a75601c5ee9482ea306c8066b6a0
--- /dev/null
+++ b/requirements/gui-dml.txt
@@ -0,0 +1,30 @@
+#1.Install torch from pytorch.org:
+#torch 2.0 with cuda 11.8
+#pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
+#torch 1.11.0 with cuda 11.3
+#pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
+einops
+fairseq
+flask
+flask_cors
+gin
+gin_config
+librosa
+local_attention
+matplotlib
+praat-parselmouth
+pyworld
+PyYAML
+resampy
+scikit_learn
+scipy
+SoundFile
+tensorboard
+tqdm
+wave
+FreeSimpleGUI
+sounddevice
+gradio
+noisereduce
+onnxruntime-directml
+torchfcpe
\ No newline at end of file
diff --git a/requirements/gui.txt b/requirements/gui.txt
new file mode 100644
index 0000000000000000000000000000000000000000..55e7bd96e527cdad5666c7bb654387f219967bdb
--- /dev/null
+++ b/requirements/gui.txt
@@ -0,0 +1,29 @@
+#1.Install torch from pytorch.org:
+#torch 2.0 with cuda 11.8
+#pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
+#torch 1.11.0 with cuda 11.3
+#pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
+einops
+fairseq
+flask
+flask_cors
+gin
+gin_config
+librosa
+local_attention
+matplotlib
+praat-parselmouth
+pyworld
+PyYAML
+resampy
+scikit_learn
+scipy
+SoundFile
+tensorboard
+tqdm
+wave
+FreeSimpleGUI
+sounddevice
+gradio
+noisereduce
+torchfcpe
diff --git a/requirements/ipex.txt b/requirements/ipex.txt
new file mode 100644
index 0000000000000000000000000000000000000000..72724f8792c3277b57a7e7418947c7ae5db4701e
--- /dev/null
+++ b/requirements/ipex.txt
@@ -0,0 +1,54 @@
+torch==2.0.1a0
+intel_extension_for_pytorch==2.0.110+xpu
+torchvision==0.15.2a0
+https://github.com/Disty0/Retrieval-based-Voice-Conversion-WebUI/releases/download/torchaudio_wheels_for_ipex/torchaudio-2.0.2+31de77d-cp310-cp310-linux_x86_64.whl
+--extra-index-url https://pytorch-extension.intel.com/release-whl/stable/xpu/us/
+joblib>=1.1.0
+numba==0.56.4
+numpy==1.23.5
+scipy
+librosa==0.9.1
+llvmlite==0.39.0
+fairseq==0.12.2
+faiss-cpu==1.7.3
+gradio
+Cython
+pydub>=0.25.1
+soundfile>=0.12.1
+tensorboardX
+Jinja2>=3.1.2
+json5
+Markdown
+matplotlib>=3.7.0
+matplotlib-inline>=0.1.3
+praat-parselmouth>=0.4.2
+Pillow>=9.1.1
+resampy>=0.4.2
+scikit-learn
+tensorboard
+tqdm>=4.63.1
+tornado>=6.1
+Werkzeug>=2.2.3
+uc-micro-py>=1.0.1
+sympy>=1.11.1
+tabulate>=0.8.10
+PyYAML>=6.0
+pyasn1>=0.4.8
+pyasn1-modules>=0.2.8
+fsspec>=2022.11.0
+absl-py>=1.2.0
+audioread
+uvicorn>=0.21.1
+colorama>=0.4.5
+pyworld==0.3.2
+httpx
+onnxruntime; sys_platform == 'darwin'
+onnxruntime-gpu; sys_platform != 'darwin'
+torchcrepe==0.0.20
+fastapi
+python-dotenv>=1.0.0
+av
+FreeSimpleGUI
+sounddevice
+torchfcpe
+pybase16384
diff --git a/requirements/main.txt b/requirements/main.txt
new file mode 100644
index 0000000000000000000000000000000000000000..2ef0eef31197f89d9220fb55ac64e552287a13cf
--- /dev/null
+++ b/requirements/main.txt
@@ -0,0 +1,47 @@
+joblib>=1.1.0
+numba
+numpy==1.23.5
+scipy
+librosa==0.9.1
+llvmlite
+fairseq
+faiss-cpu
+gradio
+Cython
+pydub>=0.25.1
+soundfile>=0.12.1
+tensorboardX
+Jinja2>=3.1.2
+json5
+Markdown
+matplotlib>=3.7.0
+matplotlib-inline>=0.1.3
+praat-parselmouth>=0.4.2
+Pillow>=9.1.1
+resampy>=0.4.2
+scikit-learn
+tensorboard
+tqdm>=4.63.1
+tornado>=6.1
+Werkzeug>=2.2.3
+uc-micro-py>=1.0.1
+sympy>=1.11.1
+tabulate>=0.8.10
+PyYAML>=6.0
+pyasn1>=0.4.8
+pyasn1-modules>=0.2.8
+fsspec>=2022.11.0
+absl-py>=1.2.0
+audioread
+uvicorn>=0.21.1
+colorama>=0.4.5
+pyworld==0.3.2
+httpx
+onnxruntime; sys_platform == 'darwin'
+onnxruntime-gpu; sys_platform != 'darwin'
+torchcrepe==0.0.20
+fastapi
+torchfcpe
+python-dotenv>=1.0.0
+av
+pybase16384
diff --git a/requirements/py311.txt b/requirements/py311.txt
new file mode 100644
index 0000000000000000000000000000000000000000..59c6b88f7bf553736b0100b09fd839e4d8db6b72
--- /dev/null
+++ b/requirements/py311.txt
@@ -0,0 +1,47 @@
+joblib>=1.1.0
+numba
+numpy
+scipy
+librosa==0.9.1
+llvmlite
+fairseq @ git+https://github.com/One-sixth/fairseq.git
+faiss-cpu
+gradio
+Cython
+pydub>=0.25.1
+soundfile>=0.12.1
+tensorboardX
+Jinja2>=3.1.2
+json5
+Markdown
+matplotlib>=3.7.0
+matplotlib-inline>=0.1.3
+praat-parselmouth>=0.4.2
+Pillow>=9.1.1
+resampy>=0.4.2
+scikit-learn
+tensorboard
+tqdm>=4.63.1
+tornado>=6.1
+Werkzeug>=2.2.3
+uc-micro-py>=1.0.1
+sympy>=1.11.1
+tabulate>=0.8.10
+PyYAML>=6.0
+pyasn1>=0.4.8
+pyasn1-modules>=0.2.8
+fsspec>=2022.11.0
+absl-py>=1.2.0
+audioread
+uvicorn>=0.21.1
+colorama>=0.4.5
+pyworld==0.3.2
+httpx
+onnxruntime; sys_platform == 'darwin'
+onnxruntime-gpu; sys_platform != 'darwin'
+torchcrepe==0.0.20
+fastapi
+torchfcpe
+python-dotenv>=1.0.0
+av
+pybase16384
diff --git a/run.sh b/run.sh
new file mode 100644
index 0000000000000000000000000000000000000000..f011baf99aaf9733fb444ce5bf9ce72359174f88
--- /dev/null
+++ b/run.sh
@@ -0,0 +1,27 @@
+#!/bin/sh
+
+set -fa
+
+# Check if Python is installed
+if ! command -v python; then
+ echo "Python not found. Please install Python using your package manager or via PyEnv."
+ exit 1
+fi
+
+requirements_file="requirements/main.txt"
+venv_path=".venv"
+
+if [[ ! -d "${venv_path}" ]]; then
+ echo "Creating venv..."
+
+ python -m venv "${venv_path}"
+ source "${venv_path}/bin/activate"
+
+ # Check if required packages are up-to-date
+ pip install --upgrade -r "${requirements_file}"
+fi
+echo "Activating venv..."
+source "${venv_path}/bin/activate"
+
+# Run the main script
+python web.py --pycmd python
diff --git a/rvc/__init__.py b/rvc/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..a7a2950c25063a198cc10c987915f7f4be49fe99
--- /dev/null
+++ b/rvc/__init__.py
@@ -0,0 +1,4 @@
+from . import ipex
+import sys
+
+del sys.modules["rvc.ipex"]
diff --git a/rvc/f0/__init__.py b/rvc/f0/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..dc91d96c9d50da0e5e32b5e1e95a730d06bb6b4a
--- /dev/null
+++ b/rvc/f0/__init__.py
@@ -0,0 +1,10 @@
+from .f0 import F0Predictor
+
+from .crepe import CRePE
+from .dio import Dio
+from .fcpe import FCPE
+from .harvest import Harvest
+from .pm import PM
+from .rmvpe import RMVPE
+
+__all__ = ["F0Predictor", "CRePE", "Dio", "FCPE", "Harvest", "PM", "RMVPE"]
diff --git a/rvc/f0/crepe.py b/rvc/f0/crepe.py
new file mode 100644
index 0000000000000000000000000000000000000000..28b8ae7d94acbdf329b7a88ae989fb8752748b10
--- /dev/null
+++ b/rvc/f0/crepe.py
@@ -0,0 +1,56 @@
+from typing import Any, Optional, Union
+
+import numpy as np
+import torch
+import torchcrepe
+
+from .f0 import F0Predictor
+
+
+class CRePE(F0Predictor):
+ def __init__(
+ self,
+ hop_length=512,
+ f0_min=50,
+ f0_max=1100,
+ sampling_rate=44100,
+ device="cpu",
+ ):
+ if "privateuseone" in str(device):
+ device = "cpu"
+ super().__init__(
+ hop_length,
+ f0_min,
+ f0_max,
+ sampling_rate,
+ device,
+ )
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = None,
+ ):
+ if p_len is None:
+ p_len = wav.shape[0] // self.hop_length
+ if not torch.is_tensor(wav):
+ wav = torch.from_numpy(wav)
+ # Pick a batch size that doesn't cause memory errors on your gpu
+ batch_size = 512
+ # Compute pitch using device 'device'
+ f0, pd = torchcrepe.predict(
+ wav.float().to(self.device).unsqueeze(dim=0),
+ self.sampling_rate,
+ self.hop_length,
+ self.f0_min,
+ self.f0_max,
+ batch_size=batch_size,
+ device=self.device,
+ return_periodicity=True,
+ )
+ pd = torchcrepe.filter.median(pd, 3)
+ f0 = torchcrepe.filter.mean(f0, 3)
+ f0[pd < 0.1] = 0
+ f0 = f0[0].cpu().numpy()
+ return self._interpolate_f0(self._resize_f0(f0, p_len))[0]
diff --git a/rvc/f0/deepunet.py b/rvc/f0/deepunet.py
new file mode 100644
index 0000000000000000000000000000000000000000..1a7517884557f471e9b00e1c3659bcf7fe98d4cb
--- /dev/null
+++ b/rvc/f0/deepunet.py
@@ -0,0 +1,217 @@
+from typing import List, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+
+class ConvBlockRes(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ momentum: float = 0.01,
+ ):
+ super(ConvBlockRes, self).__init__()
+ self.conv = nn.Sequential(
+ nn.Conv2d(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ kernel_size=(3, 3),
+ stride=(1, 1),
+ padding=(1, 1),
+ bias=False,
+ ),
+ nn.BatchNorm2d(out_channels, momentum=momentum),
+ nn.ReLU(),
+ nn.Conv2d(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ kernel_size=(3, 3),
+ stride=(1, 1),
+ padding=(1, 1),
+ bias=False,
+ ),
+ nn.BatchNorm2d(out_channels, momentum=momentum),
+ nn.ReLU(),
+ )
+ # self.shortcut:Optional[nn.Module] = None
+ if in_channels != out_channels:
+ self.shortcut = nn.Conv2d(in_channels, out_channels, (1, 1))
+
+ def forward(self, x: torch.Tensor):
+ if not hasattr(self, "shortcut"):
+ return self.conv(x) + x
+ else:
+ return self.conv(x) + self.shortcut(x)
+
+
+class Encoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ in_size: int,
+ n_encoders: int,
+ kernel_size: Tuple[int, int],
+ n_blocks: int,
+ out_channels=16,
+ momentum=0.01,
+ ):
+ super(Encoder, self).__init__()
+ self.n_encoders = n_encoders
+
+ self.bn = nn.BatchNorm2d(in_channels, momentum=momentum)
+ self.layers = nn.ModuleList()
+ for _ in range(self.n_encoders):
+ self.layers.append(
+ ResEncoderBlock(
+ in_channels, out_channels, kernel_size, n_blocks, momentum=momentum
+ )
+ )
+ in_channels = out_channels
+ out_channels *= 2
+ in_size //= 2
+ self.out_size = in_size
+ self.out_channel = out_channels
+
+ def __call__(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ return super().__call__(x)
+
+ def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ concat_tensors: List[torch.Tensor] = []
+ x = self.bn(x)
+ for layer in self.layers:
+ t, x = layer(x)
+ concat_tensors.append(t)
+ return x, concat_tensors
+
+
+class ResEncoderBlock(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ kernel_size: Tuple[int, int],
+ n_blocks=1,
+ momentum=0.01,
+ ):
+ super(ResEncoderBlock, self).__init__()
+ self.n_blocks = n_blocks
+ self.kernel_size = kernel_size
+
+ self.conv = nn.ModuleList()
+ self.conv.append(ConvBlockRes(in_channels, out_channels, momentum))
+ for _ in range(n_blocks - 1):
+ self.conv.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+ if self.kernel_size is not None:
+ self.pool = nn.AvgPool2d(kernel_size=kernel_size)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+ for conv in self.conv:
+ x = conv(x)
+ if self.kernel_size is not None:
+ return x, self.pool(x)
+ return x
+
+
+class Intermediate(nn.Module):
+ def __init__(self, in_channels, out_channels, n_inters, n_blocks, momentum=0.01):
+ super(Intermediate, self).__init__()
+
+ self.layers = nn.ModuleList()
+ self.layers.append(
+ ResEncoderBlock(in_channels, out_channels, None, n_blocks, momentum)
+ )
+ for _ in range(n_inters - 1):
+ self.layers.append(
+ ResEncoderBlock(out_channels, out_channels, None, n_blocks, momentum)
+ )
+
+ def forward(self, x):
+ for layer in self.layers:
+ x = layer(x)
+ return x
+
+
+class ResDecoderBlock(nn.Module):
+ def __init__(self, in_channels, out_channels, stride, n_blocks=1, momentum=0.01):
+ super(ResDecoderBlock, self).__init__()
+ out_padding = (0, 1) if stride == (1, 2) else (1, 1)
+
+ self.conv1 = nn.Sequential(
+ nn.ConvTranspose2d(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ kernel_size=(3, 3),
+ stride=stride,
+ padding=(1, 1),
+ output_padding=out_padding,
+ bias=False,
+ ),
+ nn.BatchNorm2d(out_channels, momentum=momentum),
+ nn.ReLU(),
+ )
+ self.conv2 = nn.ModuleList()
+ self.conv2.append(ConvBlockRes(out_channels * 2, out_channels, momentum))
+ for _ in range(n_blocks - 1):
+ self.conv2.append(ConvBlockRes(out_channels, out_channels, momentum))
+
+ def forward(self, x, concat_tensor):
+ x = self.conv1(x)
+ x = torch.cat((x, concat_tensor), dim=1)
+ for conv2 in self.conv2:
+ x = conv2(x)
+ return x
+
+
+class Decoder(nn.Module):
+ def __init__(self, in_channels, n_decoders, stride, n_blocks, momentum=0.01):
+ super(Decoder, self).__init__()
+
+ self.layers = nn.ModuleList()
+ self.n_decoders = n_decoders
+ for _ in range(self.n_decoders):
+ out_channels = in_channels // 2
+ self.layers.append(
+ ResDecoderBlock(in_channels, out_channels, stride, n_blocks, momentum)
+ )
+ in_channels = out_channels
+
+ def forward(self, x: torch.Tensor, concat_tensors: List[torch.Tensor]):
+ for i, layer in enumerate(self.layers):
+ x = layer(x, concat_tensors[-1 - i])
+ return x
+
+
+class DeepUnet(nn.Module):
+ def __init__(
+ self,
+ kernel_size: Tuple[int, int],
+ n_blocks: int,
+ en_de_layers=5,
+ inter_layers=4,
+ in_channels=1,
+ en_out_channels=16,
+ ):
+ super(DeepUnet, self).__init__()
+ self.encoder = Encoder(
+ in_channels, 128, en_de_layers, kernel_size, n_blocks, en_out_channels
+ )
+ self.intermediate = Intermediate(
+ self.encoder.out_channel // 2,
+ self.encoder.out_channel,
+ inter_layers,
+ n_blocks,
+ )
+ self.decoder = Decoder(
+ self.encoder.out_channel, en_de_layers, kernel_size, n_blocks
+ )
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ x, concat_tensors = self.encoder(x)
+ x = self.intermediate(x)
+ x = self.decoder(x, concat_tensors)
+ return x
diff --git a/rvc/f0/dio.py b/rvc/f0/dio.py
new file mode 100644
index 0000000000000000000000000000000000000000..91666cc124e19eecc43388f0ef8c4abfaa53854b
--- /dev/null
+++ b/rvc/f0/dio.py
@@ -0,0 +1,31 @@
+from typing import Any, Optional, Union
+
+import numpy as np
+import pyworld
+
+from .f0 import F0Predictor
+
+
+class Dio(F0Predictor):
+ def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+ super().__init__(hop_length, f0_min, f0_max, sampling_rate)
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = None,
+ ):
+ if p_len is None:
+ p_len = wav.shape[0] // self.hop_length
+ f0, t = pyworld.dio(
+ wav.astype(np.double),
+ fs=self.sampling_rate,
+ f0_floor=self.f0_min,
+ f0_ceil=self.f0_max,
+ frame_period=1000 * self.hop_length / self.sampling_rate,
+ )
+ f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+ for index, pitch in enumerate(f0):
+ f0[index] = round(pitch, 1)
+ return self._interpolate_f0(self._resize_f0(f0, p_len))[0]
diff --git a/rvc/f0/e2e.py b/rvc/f0/e2e.py
new file mode 100644
index 0000000000000000000000000000000000000000..5070a26a1c127c6802208918a17486eb2a8630dc
--- /dev/null
+++ b/rvc/f0/e2e.py
@@ -0,0 +1,67 @@
+from typing import Tuple
+
+import torch.nn as nn
+
+from .deepunet import DeepUnet
+
+
+class E2E(nn.Module):
+ def __init__(
+ self,
+ n_blocks: int,
+ n_gru: int,
+ kernel_size: Tuple[int, int],
+ en_de_layers=5,
+ inter_layers=4,
+ in_channels=1,
+ en_out_channels=16,
+ ):
+ super(E2E, self).__init__()
+
+ self.unet = DeepUnet(
+ kernel_size,
+ n_blocks,
+ en_de_layers,
+ inter_layers,
+ in_channels,
+ en_out_channels,
+ )
+ self.cnn = nn.Conv2d(en_out_channels, 3, (3, 3), padding=(1, 1))
+ if n_gru:
+ self.fc = nn.Sequential(
+ self.BiGRU(3 * 128, 256, n_gru),
+ nn.Linear(512, 360),
+ nn.Dropout(0.25),
+ nn.Sigmoid(),
+ )
+ else:
+ self.fc = nn.Sequential(
+ nn.Linear(3 * nn.N_MELS, nn.N_CLASS),
+ nn.Dropout(0.25),
+ nn.Sigmoid(),
+ )
+
+ def forward(self, mel):
+ mel = mel.transpose(-1, -2).unsqueeze(1)
+ x = self.cnn(self.unet(mel)).transpose(1, 2).flatten(-2)
+ x = self.fc(x)
+ return x
+
+ class BiGRU(nn.Module):
+ def __init__(
+ self,
+ input_features: int,
+ hidden_features: int,
+ num_layers: int,
+ ):
+ super().__init__()
+ self.gru = nn.GRU(
+ input_features,
+ hidden_features,
+ num_layers=num_layers,
+ batch_first=True,
+ bidirectional=True,
+ )
+
+ def forward(self, x):
+ return self.gru(x)[0]
diff --git a/rvc/f0/f0.py b/rvc/f0/f0.py
new file mode 100644
index 0000000000000000000000000000000000000000..0f615e65fa6ef723d38a9785a6084b7ce9d7f8b1
--- /dev/null
+++ b/rvc/f0/f0.py
@@ -0,0 +1,78 @@
+from typing import Optional, Union
+
+import torch
+import numpy as np
+
+
+class F0Predictor(object):
+ def __init__(
+ self,
+ hop_length=512,
+ f0_min=50,
+ f0_max=1100,
+ sampling_rate=44100,
+ device: Optional[str] = None,
+ ):
+ self.hop_length = hop_length
+ self.f0_min = f0_min
+ self.f0_max = f0_max
+ self.sampling_rate = sampling_rate
+ if device is None:
+ device = "cuda:0" if torch.cuda.is_available() else "cpu"
+ self.device = device
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = None,
+ ): ...
+
+ def _interpolate_f0(self, f0: np.ndarray):
+ """
+ 对F0进行插值处理
+ """
+
+ data = np.reshape(f0, (f0.size, 1))
+
+ vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
+ vuv_vector[data > 0.0] = 1.0
+ vuv_vector[data <= 0.0] = 0.0
+
+ ip_data = data
+
+ frame_number = data.size
+ last_value = 0.0
+ for i in range(frame_number):
+ if data[i] <= 0.0:
+ j = i + 1
+ for j in range(i + 1, frame_number):
+ if data[j] > 0.0:
+ break
+ if j < frame_number - 1:
+ if last_value > 0.0:
+ step = (data[j] - data[i - 1]) / float(j - i)
+ for k in range(i, j):
+ ip_data[k] = data[i - 1] + step * (k - i + 1)
+ else:
+ for k in range(i, j):
+ ip_data[k] = data[j]
+ else:
+ for k in range(i, frame_number):
+ ip_data[k] = last_value
+ else:
+ ip_data[i] = data[i] # 这里可能存在一个没有必要的拷贝
+ last_value = data[i]
+
+ return ip_data[:, 0], vuv_vector[:, 0]
+
+ def _resize_f0(self, x: np.ndarray, target_len: int):
+ source = np.array(x)
+ source[source < 0.001] = np.nan
+ target = np.interp(
+ np.arange(0, len(source) * target_len, len(source)) / target_len,
+ np.arange(0, len(source)),
+ source,
+ )
+ res = np.nan_to_num(target)
+ return res
diff --git a/rvc/f0/fcpe.py b/rvc/f0/fcpe.py
new file mode 100644
index 0000000000000000000000000000000000000000..d0d9410105fd6ef8d9f0a1e18fd3c3b2eae768f6
--- /dev/null
+++ b/rvc/f0/fcpe.py
@@ -0,0 +1,53 @@
+from typing import Optional, Union
+
+import numpy as np
+import torch
+
+from .f0 import F0Predictor
+
+
+class FCPE(F0Predictor):
+ def __init__(
+ self,
+ hop_length=512,
+ f0_min=50,
+ f0_max=1100,
+ sampling_rate=44100,
+ device="cpu",
+ ):
+ super().__init__(
+ hop_length,
+ f0_min,
+ f0_max,
+ sampling_rate,
+ device,
+ )
+
+ from torchfcpe import (
+ spawn_bundled_infer_model,
+ ) # must be imported at here, or it will cause fairseq crash on training
+
+ self.model = spawn_bundled_infer_model(self.device)
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = 0.006,
+ ):
+ if p_len is None:
+ p_len = wav.shape[0] // self.hop_length
+ if not torch.is_tensor(wav):
+ wav = torch.from_numpy(wav)
+ f0 = (
+ self.model.infer(
+ wav.float().to(self.device).unsqueeze(0),
+ sr=self.sampling_rate,
+ decoder_mode="local_argmax",
+ threshold=filter_radius,
+ )
+ .squeeze()
+ .cpu()
+ .numpy()
+ )
+ return self._interpolate_f0(self._resize_f0(f0, p_len))[0]
diff --git a/rvc/f0/harvest.py b/rvc/f0/harvest.py
new file mode 100644
index 0000000000000000000000000000000000000000..823c564c740fdf8ba0b5f614587edf5d9f766859
--- /dev/null
+++ b/rvc/f0/harvest.py
@@ -0,0 +1,32 @@
+from typing import Any, Optional, Union
+
+import numpy as np
+import pyworld
+from scipy import signal
+
+from .f0 import F0Predictor
+
+
+class Harvest(F0Predictor):
+ def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+ super().__init__(hop_length, f0_min, f0_max, sampling_rate)
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = None,
+ ):
+ if p_len is None:
+ p_len = wav.shape[0] // self.hop_length
+ f0, t = pyworld.harvest(
+ wav.astype(np.double),
+ fs=self.sampling_rate,
+ f0_ceil=self.f0_max,
+ f0_floor=self.f0_min,
+ frame_period=1000 * self.hop_length / self.sampling_rate,
+ )
+ f0 = pyworld.stonemask(wav.astype(np.double), f0, t, self.sampling_rate)
+ if filter_radius is not None and filter_radius > 2:
+ f0 = signal.medfilt(f0, filter_radius)
+ return self._interpolate_f0(self._resize_f0(f0, p_len))[0]
diff --git a/rvc/f0/mel.py b/rvc/f0/mel.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c068002e444266d07a85b7266c5df17e940a53c
--- /dev/null
+++ b/rvc/f0/mel.py
@@ -0,0 +1,71 @@
+from typing import Optional
+
+import torch
+import numpy as np
+from librosa.filters import mel
+
+from .stft import STFT
+
+
+class MelSpectrogram(torch.nn.Module):
+ def __init__(
+ self,
+ is_half: bool,
+ n_mel_channels: int,
+ sampling_rate: int,
+ win_length: int,
+ hop_length: int,
+ n_fft: Optional[int] = None,
+ mel_fmin: int = 0,
+ mel_fmax: int = None,
+ clamp: float = 1e-5,
+ device=torch.device("cpu"),
+ ):
+ super().__init__()
+ if n_fft is None:
+ n_fft = win_length
+ mel_basis = mel(
+ sr=sampling_rate,
+ n_fft=n_fft,
+ n_mels=n_mel_channels,
+ fmin=mel_fmin,
+ fmax=mel_fmax,
+ htk=True,
+ )
+ mel_basis = torch.from_numpy(mel_basis).float()
+ self.register_buffer("mel_basis", mel_basis)
+ self.n_fft = n_fft
+ self.hop_length = hop_length
+ self.win_length = win_length
+ self.clamp = clamp
+ self.is_half = is_half
+
+ self.stft = STFT(
+ filter_length=n_fft,
+ hop_length=hop_length,
+ win_length=win_length,
+ window="hann",
+ use_torch_stft="privateuseone" not in str(device),
+ ).to(device)
+
+ def forward(
+ self,
+ audio: torch.Tensor,
+ keyshift=0,
+ speed=1,
+ center=True,
+ ):
+ factor = 2 ** (keyshift / 12)
+ win_length_new = int(np.round(self.win_length * factor))
+ magnitude = self.stft(audio, keyshift, speed, center)
+ if keyshift != 0:
+ size = self.n_fft // 2 + 1
+ resize = magnitude.size(1)
+ if resize < size:
+ magnitude = torch.nn.functional.pad(magnitude, (0, 0, 0, size - resize))
+ magnitude = magnitude[:, :size, :] * self.win_length / win_length_new
+ mel_output = torch.matmul(self.mel_basis, magnitude)
+ if self.is_half:
+ mel_output = mel_output.half()
+ log_mel_spec = torch.log(torch.clamp(mel_output, min=self.clamp))
+ return log_mel_spec
diff --git a/rvc/f0/models.py b/rvc/f0/models.py
new file mode 100644
index 0000000000000000000000000000000000000000..7c2853e04b51b63e76ef8ee7d662159b28136eb1
--- /dev/null
+++ b/rvc/f0/models.py
@@ -0,0 +1,16 @@
+import torch
+
+
+def get_rmvpe(
+ model_path="assets/rmvpe/rmvpe.pt", device=torch.device("cpu"), is_half=False
+):
+ from rvc.f0.e2e import E2E
+
+ model = E2E(4, 1, (2, 2))
+ ckpt = torch.load(model_path, map_location=device)
+ model.load_state_dict(ckpt)
+ model.eval()
+ if is_half:
+ model = model.half()
+ model = model.to(device)
+ return model
diff --git a/rvc/f0/pm.py b/rvc/f0/pm.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf54b3c0707586361df0edb0c1a9cb4150448305
--- /dev/null
+++ b/rvc/f0/pm.py
@@ -0,0 +1,39 @@
+from typing import Any, Optional
+
+import numpy as np
+import parselmouth
+
+from .f0 import F0Predictor
+
+
+class PM(F0Predictor):
+ def __init__(self, hop_length=512, f0_min=50, f0_max=1100, sampling_rate=44100):
+ super().__init__(hop_length, f0_min, f0_max, sampling_rate)
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[int] = None,
+ ):
+ x = wav
+ if p_len is None:
+ p_len = x.shape[0] // self.hop_length
+ else:
+ assert abs(p_len - x.shape[0] // self.hop_length) < 4, "pad length error"
+ time_step = self.hop_length / self.sampling_rate * 1000
+ f0 = (
+ parselmouth.Sound(x, self.sampling_rate)
+ .to_pitch_ac(
+ time_step=time_step / 1000,
+ voicing_threshold=0.6,
+ pitch_floor=self.f0_min,
+ pitch_ceiling=self.f0_max,
+ )
+ .selected_array["frequency"]
+ )
+
+ pad_size = (p_len - len(f0) + 1) // 2
+ if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+ f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+ return self._interpolate_f0(f0)[0]
diff --git a/rvc/f0/rmvpe.py b/rvc/f0/rmvpe.py
new file mode 100644
index 0000000000000000000000000000000000000000..b1de448783515bfd19c5730918535f35acebf1cc
--- /dev/null
+++ b/rvc/f0/rmvpe.py
@@ -0,0 +1,162 @@
+from io import BytesIO
+import os
+from typing import Any, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from rvc.jit import load_inputs, get_jit_model, export_jit_model, save_pickle
+
+from .mel import MelSpectrogram
+from .f0 import F0Predictor
+from .models import get_rmvpe
+
+
+def rmvpe_jit_export(
+ model_path: str,
+ mode: str = "script",
+ inputs_path: str = None,
+ save_path: str = None,
+ device=torch.device("cpu"),
+ is_half=False,
+):
+ if not save_path:
+ save_path = model_path.rstrip(".pth")
+ save_path += ".half.jit" if is_half else ".jit"
+ if "cuda" in str(device) and ":" not in str(device):
+ device = torch.device("cuda:0")
+
+ model = get_rmvpe(model_path, device, is_half)
+ inputs = None
+ if mode == "trace":
+ inputs = load_inputs(inputs_path, device, is_half)
+ ckpt = export_jit_model(model, mode, inputs, device, is_half)
+ ckpt["device"] = str(device)
+ save_pickle(ckpt, save_path)
+ return ckpt
+
+
+class RMVPE(F0Predictor):
+ def __init__(
+ self,
+ model_path: str,
+ is_half: bool,
+ device: str,
+ use_jit=False,
+ ):
+ hop_length = 160
+ f0_min = 30
+ f0_max = 8000
+ sampling_rate = 16000
+
+ super().__init__(
+ hop_length,
+ f0_min,
+ f0_max,
+ sampling_rate,
+ device,
+ )
+
+ self.is_half = is_half
+ cents_mapping = 20 * np.arange(360) + 1997.3794084376191
+ self.cents_mapping = np.pad(cents_mapping, (4, 4)) # 368
+
+ self.mel_extractor = MelSpectrogram(
+ is_half=is_half,
+ n_mel_channels=128,
+ sampling_rate=sampling_rate,
+ win_length=1024,
+ hop_length=hop_length,
+ mel_fmin=f0_min,
+ mel_fmax=f0_max,
+ device=self.device,
+ ).to(self.device)
+
+ if "privateuseone" in str(self.device):
+ import onnxruntime as ort
+
+ self.model = ort.InferenceSession(
+ "%s/rmvpe.onnx" % os.environ["rmvpe_root"],
+ providers=["DmlExecutionProvider"],
+ )
+ else:
+
+ def rmvpe_jit_model():
+ ckpt = get_jit_model(model_path, is_half, self.device, rmvpe_jit_export)
+ model = torch.jit.load(BytesIO(ckpt["model"]), map_location=self.device)
+ model = model.to(self.device)
+ return model
+
+ if use_jit and not (is_half and "cpu" in str(self.device)):
+ self.model = rmvpe_jit_model()
+ else:
+ self.model = get_rmvpe(model_path, self.device, is_half)
+
+ def compute_f0(
+ self,
+ wav: np.ndarray,
+ p_len: Optional[int] = None,
+ filter_radius: Optional[Union[int, float]] = None,
+ ):
+ if p_len is None:
+ p_len = wav.shape[0] // self.hop_length
+ if not torch.is_tensor(wav):
+ wav = torch.from_numpy(wav)
+ mel = self.mel_extractor(wav.float().to(self.device).unsqueeze(0), center=True)
+ hidden = self._mel2hidden(mel)
+ if "privateuseone" not in str(self.device):
+ hidden = hidden.squeeze(0).cpu().numpy()
+ else:
+ hidden = hidden[0]
+ if self.is_half == True:
+ hidden = hidden.astype("float32")
+
+ f0 = self._decode(hidden, thred=filter_radius)
+
+ return self._interpolate_f0(self._resize_f0(f0, p_len))[0]
+
+ def _to_local_average_cents(self, salience, threshold=0.05):
+ center = np.argmax(salience, axis=1) # 帧长#index
+ salience = np.pad(salience, ((0, 0), (4, 4))) # 帧长,368
+ center += 4
+ todo_salience = []
+ todo_cents_mapping = []
+ starts = center - 4
+ ends = center + 5
+ for idx in range(salience.shape[0]):
+ todo_salience.append(salience[:, starts[idx] : ends[idx]][idx])
+ todo_cents_mapping.append(self.cents_mapping[starts[idx] : ends[idx]])
+ todo_salience = np.array(todo_salience) # 帧长，9
+ todo_cents_mapping = np.array(todo_cents_mapping) # 帧长，9
+ product_sum = np.sum(todo_salience * todo_cents_mapping, 1)
+ weight_sum = np.sum(todo_salience, 1) # 帧长
+ devided = product_sum / weight_sum # 帧长
+ maxx = np.max(salience, axis=1) # 帧长
+ devided[maxx <= threshold] = 0
+ return devided
+
+ def _mel2hidden(self, mel):
+ with torch.no_grad():
+ n_frames = mel.shape[-1]
+ n_pad = 32 * ((n_frames - 1) // 32 + 1) - n_frames
+ if n_pad > 0:
+ mel = F.pad(mel, (0, n_pad), mode="constant")
+ if "privateuseone" in str(self.device):
+ onnx_input_name = self.model.get_inputs()[0].name
+ onnx_outputs_names = self.model.get_outputs()[0].name
+ hidden = self.model.run(
+ [onnx_outputs_names],
+ input_feed={onnx_input_name: mel.cpu().numpy()},
+ )[0]
+ else:
+ mel = mel.half() if self.is_half else mel.float()
+ hidden = self.model(mel)
+ return hidden[:, :n_frames]
+
+ def _decode(self, hidden, thred=0.03):
+ cents_pred = self._to_local_average_cents(hidden, threshold=thred)
+ f0 = 10 * (2 ** (cents_pred / 1200))
+ f0[f0 == 10] = 0
+ # f0 = np.array([10 * (2 ** (cent_pred / 1200)) if cent_pred else 0 for cent_pred in cents_pred])
+ return f0
diff --git a/rvc/f0/stft.py b/rvc/f0/stft.py
new file mode 100644
index 0000000000000000000000000000000000000000..b84daa8867a5d154df7d2dfd40e19f1bd255fd7d
--- /dev/null
+++ b/rvc/f0/stft.py
@@ -0,0 +1,194 @@
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from librosa.util import pad_center
+from scipy.signal import get_window
+
+
+class STFT(torch.nn.Module):
+ def __init__(
+ self,
+ filter_length=1024,
+ hop_length=512,
+ win_length: Optional[int] = None,
+ window="hann",
+ use_torch_stft=True,
+ ):
+ """
+ This module implements an STFT using 1D convolution and 1D transpose convolutions.
+ This is a bit tricky so there are some cases that probably won't work as working
+ out the same sizes before and after in all overlap add setups is tough. Right now,
+ this code should work with hop lengths that are half the filter length (50% overlap
+ between frames).
+
+ Keyword Arguments:
+ filter_length {int} -- Length of filters used (default: {1024})
+ hop_length {int} -- Hop length of STFT (restrict to 50% overlap between frames) (default: {512})
+ win_length {[type]} -- Length of the window function applied to each frame (if not specified, it
+ equals the filter length). (default: {None})
+ window {str} -- Type of window to use (options are bartlett, hann, hamming, blackman, blackmanharris)
+ (default: {'hann'})
+ """
+ super(STFT, self).__init__()
+ self.filter_length = filter_length
+ self.hop_length = hop_length
+ self.pad_amount = int(self.filter_length / 2)
+ self.win_length = win_length
+ self.hann_window = {}
+ self.use_torch_stft = use_torch_stft
+
+ if use_torch_stft:
+ return
+
+ fourier_basis = np.fft.fft(np.eye(self.filter_length))
+
+ cutoff = int((self.filter_length / 2 + 1))
+ fourier_basis = np.vstack(
+ [np.real(fourier_basis[:cutoff, :]), np.imag(fourier_basis[:cutoff, :])]
+ )
+ forward_basis = torch.FloatTensor(fourier_basis)
+ inverse_basis = torch.FloatTensor(np.linalg.pinv(fourier_basis))
+
+ if win_length is None or not win_length:
+ win_length = filter_length
+ assert filter_length >= win_length
+
+ # get window and zero center pad it to filter_length
+ fft_window = get_window(window, win_length, fftbins=True)
+ fft_window = pad_center(fft_window, size=filter_length)
+ fft_window = torch.from_numpy(fft_window).float()
+
+ # window the bases
+ forward_basis *= fft_window
+ inverse_basis = (inverse_basis.T * fft_window).T
+
+ self.register_buffer("forward_basis", forward_basis.float())
+ self.register_buffer("inverse_basis", inverse_basis.float())
+ self.register_buffer("fft_window", fft_window.float())
+
+ def __call__(
+ self,
+ input_data: torch.Tensor,
+ keyshift: int = 0,
+ speed: int = 1,
+ center: bool = True,
+ ) -> torch.Tensor:
+ return super().__call__(input_data, keyshift, speed, center)
+
+ def transform(
+ self,
+ input_data: torch.Tensor,
+ return_phase=False,
+ ) -> Tuple[Union[Tuple[torch.Tensor, torch.Tensor], torch.Tensor]]:
+ """Take input data (audio) to STFT domain.
+
+ Arguments:
+ input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+ Returns:
+ magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+ num_frequencies, num_frames)
+ phase {tensor} -- Phase of STFT with shape (num_batch,
+ num_frequencies, num_frames)
+ """
+ input_data = F.pad(
+ input_data,
+ (self.pad_amount, self.pad_amount),
+ mode="reflect",
+ )
+ forward_transform = input_data.unfold(
+ 1, self.filter_length, self.hop_length
+ ).permute(0, 2, 1)
+ forward_transform = torch.matmul(self.forward_basis, forward_transform)
+ cutoff = int((self.filter_length / 2) + 1)
+ real_part = forward_transform[:, :cutoff, :]
+ imag_part = forward_transform[:, cutoff:, :]
+ magnitude = torch.sqrt(real_part**2 + imag_part**2)
+ if return_phase:
+ phase = torch.atan2(imag_part.data, real_part.data)
+ return magnitude, phase
+ else:
+ return magnitude
+
+ def inverse(
+ self,
+ magnitude: torch.Tensor,
+ phase: torch.Tensor,
+ ) -> torch.Tensor:
+ """Call the inverse STFT (iSTFT), given magnitude and phase tensors produced
+ by the ```transform``` function.
+
+ Arguments:
+ magnitude {tensor} -- Magnitude of STFT with shape (num_batch,
+ num_frequencies, num_frames)
+ phase {tensor} -- Phase of STFT with shape (num_batch,
+ num_frequencies, num_frames)
+
+ Returns:
+ inverse_transform {tensor} -- Reconstructed audio given magnitude and phase. Of
+ shape (num_batch, num_samples)
+ """
+ cat = torch.cat(
+ [magnitude * torch.cos(phase), magnitude * torch.sin(phase)], dim=1
+ )
+ fold = torch.nn.Fold(
+ output_size=(1, (cat.size(-1) - 1) * self.hop_length + self.filter_length),
+ kernel_size=(1, self.filter_length),
+ stride=(1, self.hop_length),
+ )
+ inverse_transform = torch.matmul(self.inverse_basis, cat)
+ inverse_transform: torch.Tensor = fold(inverse_transform)[
+ :, 0, 0, self.pad_amount : -self.pad_amount
+ ]
+ window_square_sum = (
+ self.fft_window.pow(2).repeat(cat.size(-1), 1).T.unsqueeze(0)
+ )
+ window_square_sum = fold(window_square_sum)[
+ :, 0, 0, self.pad_amount : -self.pad_amount
+ ]
+ inverse_transform /= window_square_sum
+ return inverse_transform
+
+ def forward(
+ self,
+ input_data: torch.Tensor,
+ keyshift: int = 0,
+ speed: int = 1,
+ center: bool = True,
+ ) -> torch.Tensor:
+ factor = 2 ** (keyshift / 12)
+ n_fft_new = int(np.round(self.filter_length * factor))
+ win_length_new = int(np.round(self.win_length * factor))
+ hop_length_new = int(np.round(self.hop_length * speed))
+ if self.use_torch_stft:
+ keyshift_key = str(keyshift) + "_" + str(input_data.device)
+ if keyshift_key not in self.hann_window:
+ self.hann_window[keyshift_key] = torch.hann_window(
+ self.win_length,
+ ).to(input_data.device)
+ fft = torch.stft(
+ input_data,
+ n_fft=n_fft_new,
+ hop_length=hop_length_new,
+ win_length=win_length_new,
+ window=self.hann_window[keyshift_key],
+ center=center,
+ return_complex=True,
+ )
+ return torch.sqrt(fft.real.pow(2) + fft.imag.pow(2))
+ return self.transform(input_data)
+ """Take input data (audio) to STFT domain and then back to audio.
+
+ Arguments:
+ input_data {tensor} -- Tensor of floats, with shape (num_batch, num_samples)
+
+ Returns:
+ reconstruction {tensor} -- Reconstructed audio given magnitude and phase. Of
+ shape (num_batch, num_samples)
+ reconstruction = self.inverse(
+ self.transform(input_data, return_phase=True),
+ )
+ return reconstruction
+ """
diff --git a/rvc/hubert.py b/rvc/hubert.py
new file mode 100644
index 0000000000000000000000000000000000000000..b9ac6b0b9e79b680f7a5cc8dfed72e3e7d6ac44c
--- /dev/null
+++ b/rvc/hubert.py
@@ -0,0 +1,339 @@
+import math
+import random
+from typing import Optional, Tuple
+
+from fairseq.checkpoint_utils import load_model_ensemble_and_task
+from fairseq.utils import index_put
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+
+# @torch.jit.script
+def pad_to_multiple(x, multiple, dim=-1, value=0):
+ # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41
+ if x is None:
+ return None, 0
+ tsz = x.size(dim)
+ m = tsz / multiple
+ remainder = math.ceil(m) * multiple - tsz
+ if int(tsz % multiple) == 0:
+ return x, 0
+ pad_offset = (0,) * (-1 - dim) * 2
+
+ return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder
+
+
+def extract_features(
+ self,
+ x,
+ padding_mask=None,
+ tgt_layer=None,
+ min_layer=0,
+):
+ if padding_mask is not None:
+ x = index_put(x, padding_mask, 0)
+
+ x_conv = self.pos_conv(x.transpose(1, 2))
+ x_conv = x_conv.transpose(1, 2)
+ x = x + x_conv
+
+ if not self.layer_norm_first:
+ x = self.layer_norm(x)
+
+ # pad to the sequence length dimension
+ x, pad_length = pad_to_multiple(x, self.required_seq_len_multiple, dim=-2, value=0)
+ if pad_length > 0 and padding_mask is None:
+ padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool)
+ padding_mask[:, -pad_length:] = True
+ else:
+ padding_mask, _ = pad_to_multiple(
+ padding_mask, self.required_seq_len_multiple, dim=-1, value=True
+ )
+ x = F.dropout(x, p=self.dropout, training=self.training)
+
+ # B x T x C -> T x B x C
+ x = x.transpose(0, 1)
+
+ layer_results = []
+ r = None
+ for i, layer in enumerate(self.layers):
+ dropout_probability = np.random.random() if self.layerdrop > 0 else 1
+ if not self.training or (dropout_probability > self.layerdrop):
+ x, (z, lr) = layer(
+ x, self_attn_padding_mask=padding_mask, need_weights=False
+ )
+ if i >= min_layer:
+ layer_results.append((x, z, lr))
+ if i == tgt_layer:
+ r = x
+ break
+
+ if r is not None:
+ x = r
+
+ # T x B x C -> B x T x C
+ x = x.transpose(0, 1)
+
+ # undo paddding
+ if pad_length > 0:
+ x = x[:, :-pad_length]
+
+ def undo_pad(a, b, c):
+ return (
+ a[:-pad_length],
+ b[:-pad_length] if b is not None else b,
+ c[:-pad_length],
+ )
+
+ layer_results = [undo_pad(*u) for u in layer_results]
+
+ return x, layer_results
+
+
+def compute_mask_indices(
+ shape: Tuple[int, int],
+ padding_mask: Optional[torch.Tensor],
+ mask_prob: float,
+ mask_length: int,
+ mask_type: str = "static",
+ mask_other: float = 0.0,
+ min_masks: int = 0,
+ no_overlap: bool = False,
+ min_space: int = 0,
+ require_same_masks: bool = True,
+ mask_dropout: float = 0.0,
+) -> torch.Tensor:
+ """
+ Computes random mask spans for a given shape
+
+ Args:
+ shape: the the shape for which to compute masks.
+ should be of size 2 where first element is batch size and 2nd is timesteps
+ padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+ mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+ number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+ however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+ mask_type: how to compute mask lengths
+ static = fixed size
+ uniform = sample from uniform distribution [mask_other, mask_length*2]
+ normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+ poisson = sample from possion distribution with lambda = mask length
+ min_masks: minimum number of masked spans
+ no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+ min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+ require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample
+ mask_dropout: randomly dropout this percentage of masks in each example
+ """
+
+ bsz, all_sz = shape
+ mask = torch.full((bsz, all_sz), False)
+
+ all_num_mask = int(
+ # add a random number for probabilistic rounding
+ mask_prob * all_sz / float(mask_length)
+ + torch.rand([1]).item()
+ )
+
+ all_num_mask = max(min_masks, all_num_mask)
+
+ mask_idcs = []
+ for i in range(bsz):
+ if padding_mask is not None:
+ sz = all_sz - padding_mask[i].long().sum().item()
+ num_mask = int(mask_prob * sz / float(mask_length) + np.random.rand())
+ num_mask = max(min_masks, num_mask)
+ else:
+ sz = all_sz
+ num_mask = all_num_mask
+
+ if mask_type == "static":
+ lengths = torch.full([num_mask], mask_length)
+ elif mask_type == "uniform":
+ lengths = torch.randint(mask_other, mask_length * 2 + 1, size=[num_mask])
+ elif mask_type == "normal":
+ lengths = torch.normal(mask_length, mask_other, size=[num_mask])
+ lengths = [max(1, int(round(x))) for x in lengths]
+ else:
+ raise Exception("unknown mask selection " + mask_type)
+
+ if sum(lengths) == 0:
+ lengths[0] = min(mask_length, sz - 1)
+
+ if no_overlap:
+ mask_idc = []
+
+ def arrange(s, e, length, keep_length):
+ span_start = torch.randint(low=s, high=e - length, size=[1]).item()
+ mask_idc.extend(span_start + i for i in range(length))
+
+ new_parts = []
+ if span_start - s - min_space >= keep_length:
+ new_parts.append((s, span_start - min_space + 1))
+ if e - span_start - length - min_space > keep_length:
+ new_parts.append((span_start + length + min_space, e))
+ return new_parts
+
+ parts = [(0, sz)]
+ min_length = min(lengths)
+ for length in sorted(lengths, reverse=True):
+ t = [e - s if e - s >= length + min_space else 0 for s, e in parts]
+ lens = torch.asarray(t, dtype=torch.int)
+ l_sum = torch.sum(lens)
+ if l_sum == 0:
+ break
+ probs = lens / torch.sum(lens)
+ c = torch.multinomial(probs.float(), len(parts)).item()
+ s, e = parts.pop(c)
+ parts.extend(arrange(s, e, length, min_length))
+ mask_idc = torch.asarray(mask_idc)
+ else:
+ min_len = min(lengths)
+ if sz - min_len <= num_mask:
+ min_len = sz - num_mask - 1
+ mask_idc = torch.asarray(
+ random.sample([i for i in range(sz - min_len)], num_mask)
+ )
+ mask_idc = torch.asarray(
+ [
+ mask_idc[j] + offset
+ for j in range(len(mask_idc))
+ for offset in range(lengths[j])
+ ]
+ )
+
+ mask_idcs.append(torch.unique(mask_idc[mask_idc < sz]))
+
+ min_len = min([len(m) for m in mask_idcs])
+ for i, mask_idc in enumerate(mask_idcs):
+ if isinstance(mask_idc, torch.Tensor):
+ mask_idc = torch.asarray(mask_idc, dtype=torch.float)
+ if len(mask_idc) > min_len and require_same_masks:
+ mask_idc = torch.asarray(
+ random.sample([i for i in range(mask_idc)], min_len)
+ )
+ if mask_dropout > 0:
+ num_holes = int(round(len(mask_idc) * mask_dropout))
+ mask_idc = torch.asarray(
+ random.sample([i for i in range(mask_idc)], len(mask_idc) - num_holes)
+ )
+
+ mask[i, mask_idc.int()] = True
+
+ return mask
+
+
+def apply_mask(self, x, padding_mask, target_list):
+ B, T, C = x.shape
+ torch.zeros_like(x)
+ if self.mask_prob > 0:
+ mask_indices = compute_mask_indices(
+ (B, T),
+ padding_mask,
+ self.mask_prob,
+ self.mask_length,
+ self.mask_selection,
+ self.mask_other,
+ min_masks=2,
+ no_overlap=self.no_mask_overlap,
+ min_space=self.mask_min_space,
+ )
+ mask_indices = mask_indices.to(x.device)
+ x[mask_indices] = self.mask_emb
+ else:
+ mask_indices = None
+
+ if self.mask_channel_prob > 0:
+ mask_channel_indices = compute_mask_indices(
+ (B, C),
+ None,
+ self.mask_channel_prob,
+ self.mask_channel_length,
+ self.mask_channel_selection,
+ self.mask_channel_other,
+ no_overlap=self.no_mask_channel_overlap,
+ min_space=self.mask_channel_min_space,
+ )
+ mask_channel_indices = (
+ mask_channel_indices.to(x.device).unsqueeze(1).expand(-1, T, -1)
+ )
+ x[mask_channel_indices] = 0
+
+ return x, mask_indices
+
+
+def get_hubert(model_path="assets/hubert/hubert_base.pt", device=torch.device("cpu")):
+ models, _, _ = load_model_ensemble_and_task(
+ [model_path],
+ suffix="",
+ )
+ hubert_model = models[0]
+ hubert_model = hubert_model.to(device)
+
+ def _apply_mask(x, padding_mask, target_list):
+ return apply_mask(hubert_model, x, padding_mask, target_list)
+
+ hubert_model.apply_mask = _apply_mask
+
+ def _extract_features(
+ x,
+ padding_mask=None,
+ tgt_layer=None,
+ min_layer=0,
+ ):
+ return extract_features(
+ hubert_model.encoder,
+ x,
+ padding_mask=padding_mask,
+ tgt_layer=tgt_layer,
+ min_layer=min_layer,
+ )
+
+ hubert_model.encoder.extract_features = _extract_features
+
+ hubert_model._forward = hubert_model.forward
+
+ def hubert_extract_features(
+ self,
+ source: torch.Tensor,
+ padding_mask: Optional[torch.Tensor] = None,
+ mask: bool = False,
+ ret_conv: bool = False,
+ output_layer: Optional[int] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ res = self._forward(
+ source,
+ padding_mask=padding_mask,
+ mask=mask,
+ features_only=True,
+ output_layer=output_layer,
+ )
+ feature = res["features"] if ret_conv else res["x"]
+ return feature, res["padding_mask"]
+
+ def _hubert_extract_features(
+ source: torch.Tensor,
+ padding_mask: Optional[torch.Tensor] = None,
+ mask: bool = False,
+ ret_conv: bool = False,
+ output_layer: Optional[int] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ return hubert_extract_features(
+ hubert_model, source, padding_mask, mask, ret_conv, output_layer
+ )
+
+ hubert_model.extract_features = _hubert_extract_features
+
+ def infer(source, padding_mask, output_layer: torch.Tensor):
+ output_layer = output_layer.item()
+ logits = hubert_model.extract_features(
+ source=source, padding_mask=padding_mask, output_layer=output_layer
+ )
+ feats = hubert_model.final_proj(logits[0]) if output_layer == 9 else logits[0]
+ return feats
+
+ hubert_model.infer = infer
+ # hubert_model.forward=infer
+ # hubert_model.forward
+
+ return hubert_model
diff --git a/rvc/ipex/__init__.py b/rvc/ipex/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..93f59a4218544256fc173390524a111a9a9baf8b
--- /dev/null
+++ b/rvc/ipex/__init__.py
@@ -0,0 +1,10 @@
+try:
+ import torch
+
+ if torch.xpu.is_available():
+ from .init import ipex_init
+
+ ipex_init()
+ from .gradscaler import gradscaler_init
+except Exception: # pylint: disable=broad-exception-caught
+ pass
diff --git a/rvc/ipex/attention.py b/rvc/ipex/attention.py
new file mode 100644
index 0000000000000000000000000000000000000000..78a4775ccf95ded03a953e07e5ffccc7bb4f29b5
--- /dev/null
+++ b/rvc/ipex/attention.py
@@ -0,0 +1,218 @@
+import torch
+import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+original_torch_bmm = torch.bmm
+
+
+def torch_bmm(input, mat2, *, out=None):
+ if input.dtype != mat2.dtype:
+ mat2 = mat2.to(input.dtype)
+
+ # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
+ batch_size_attention, input_tokens, mat2_shape = (
+ input.shape[0],
+ input.shape[1],
+ mat2.shape[2],
+ )
+ block_multiply = input.element_size()
+ slice_block_size = input_tokens * mat2_shape / 1024 / 1024 * block_multiply
+ block_size = batch_size_attention * slice_block_size
+
+ split_slice_size = batch_size_attention
+ if block_size > 4:
+ do_split = True
+ # Find something divisible with the input_tokens
+ while (split_slice_size * slice_block_size) > 4:
+ split_slice_size = split_slice_size // 2
+ if split_slice_size <= 1:
+ split_slice_size = 1
+ break
+ else:
+ do_split = False
+
+ split_2_slice_size = input_tokens
+ if split_slice_size * slice_block_size > 4:
+ slice_block_size2 = split_slice_size * mat2_shape / 1024 / 1024 * block_multiply
+ do_split_2 = True
+ # Find something divisible with the input_tokens
+ while (split_2_slice_size * slice_block_size2) > 4:
+ split_2_slice_size = split_2_slice_size // 2
+ if split_2_slice_size <= 1:
+ split_2_slice_size = 1
+ break
+ else:
+ do_split_2 = False
+
+ if do_split:
+ hidden_states = torch.zeros(
+ input.shape[0],
+ input.shape[1],
+ mat2.shape[2],
+ device=input.device,
+ dtype=input.dtype,
+ )
+ for i in range(batch_size_attention // split_slice_size):
+ start_idx = i * split_slice_size
+ end_idx = (i + 1) * split_slice_size
+ if do_split_2:
+ for i2 in range(
+ input_tokens // split_2_slice_size
+ ): # pylint: disable=invalid-name
+ start_idx_2 = i2 * split_2_slice_size
+ end_idx_2 = (i2 + 1) * split_2_slice_size
+ hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = (
+ original_torch_bmm(
+ input[start_idx:end_idx, start_idx_2:end_idx_2],
+ mat2[start_idx:end_idx, start_idx_2:end_idx_2],
+ out=out,
+ )
+ )
+ else:
+ hidden_states[start_idx:end_idx] = original_torch_bmm(
+ input[start_idx:end_idx], mat2[start_idx:end_idx], out=out
+ )
+ else:
+ return original_torch_bmm(input, mat2, out=out)
+ return hidden_states
+
+
+original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+
+
+def scaled_dot_product_attention(
+ query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False
+):
+ # ARC GPUs can't allocate more than 4GB to a single block, Slice it:
+ if len(query.shape) == 3:
+ batch_size_attention, query_tokens, shape_four = query.shape
+ shape_one = 1
+ no_shape_one = True
+ else:
+ shape_one, batch_size_attention, query_tokens, shape_four = query.shape
+ no_shape_one = False
+
+ block_multiply = query.element_size()
+ slice_block_size = (
+ shape_one * query_tokens * shape_four / 1024 / 1024 * block_multiply
+ )
+ block_size = batch_size_attention * slice_block_size
+
+ split_slice_size = batch_size_attention
+ if block_size > 4:
+ do_split = True
+ # Find something divisible with the shape_one
+ while (split_slice_size * slice_block_size) > 4:
+ split_slice_size = split_slice_size // 2
+ if split_slice_size <= 1:
+ split_slice_size = 1
+ break
+ else:
+ do_split = False
+
+ split_2_slice_size = query_tokens
+ if split_slice_size * slice_block_size > 4:
+ slice_block_size2 = (
+ shape_one * split_slice_size * shape_four / 1024 / 1024 * block_multiply
+ )
+ do_split_2 = True
+ # Find something divisible with the batch_size_attention
+ while (split_2_slice_size * slice_block_size2) > 4:
+ split_2_slice_size = split_2_slice_size // 2
+ if split_2_slice_size <= 1:
+ split_2_slice_size = 1
+ break
+ else:
+ do_split_2 = False
+
+ if do_split:
+ hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
+ for i in range(batch_size_attention // split_slice_size):
+ start_idx = i * split_slice_size
+ end_idx = (i + 1) * split_slice_size
+ if do_split_2:
+ for i2 in range(
+ query_tokens // split_2_slice_size
+ ): # pylint: disable=invalid-name
+ start_idx_2 = i2 * split_2_slice_size
+ end_idx_2 = (i2 + 1) * split_2_slice_size
+ if no_shape_one:
+ hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = (
+ original_scaled_dot_product_attention(
+ query[start_idx:end_idx, start_idx_2:end_idx_2],
+ key[start_idx:end_idx, start_idx_2:end_idx_2],
+ value[start_idx:end_idx, start_idx_2:end_idx_2],
+ attn_mask=(
+ attn_mask[start_idx:end_idx, start_idx_2:end_idx_2]
+ if attn_mask is not None
+ else attn_mask
+ ),
+ dropout_p=dropout_p,
+ is_causal=is_causal,
+ )
+ )
+ else:
+ hidden_states[:, start_idx:end_idx, start_idx_2:end_idx_2] = (
+ original_scaled_dot_product_attention(
+ query[:, start_idx:end_idx, start_idx_2:end_idx_2],
+ key[:, start_idx:end_idx, start_idx_2:end_idx_2],
+ value[:, start_idx:end_idx, start_idx_2:end_idx_2],
+ attn_mask=(
+ attn_mask[
+ :, start_idx:end_idx, start_idx_2:end_idx_2
+ ]
+ if attn_mask is not None
+ else attn_mask
+ ),
+ dropout_p=dropout_p,
+ is_causal=is_causal,
+ )
+ )
+ else:
+ if no_shape_one:
+ hidden_states[start_idx:end_idx] = (
+ original_scaled_dot_product_attention(
+ query[start_idx:end_idx],
+ key[start_idx:end_idx],
+ value[start_idx:end_idx],
+ attn_mask=(
+ attn_mask[start_idx:end_idx]
+ if attn_mask is not None
+ else attn_mask
+ ),
+ dropout_p=dropout_p,
+ is_causal=is_causal,
+ )
+ )
+ else:
+ hidden_states[:, start_idx:end_idx] = (
+ original_scaled_dot_product_attention(
+ query[:, start_idx:end_idx],
+ key[:, start_idx:end_idx],
+ value[:, start_idx:end_idx],
+ attn_mask=(
+ attn_mask[:, start_idx:end_idx]
+ if attn_mask is not None
+ else attn_mask
+ ),
+ dropout_p=dropout_p,
+ is_causal=is_causal,
+ )
+ )
+ else:
+ return original_scaled_dot_product_attention(
+ query,
+ key,
+ value,
+ attn_mask=attn_mask,
+ dropout_p=dropout_p,
+ is_causal=is_causal,
+ )
+ return hidden_states
+
+
+def attention_init():
+ # ARC GPUs can't allocate more than 4GB to a single block:
+ torch.bmm = torch_bmm
+ torch.nn.functional.scaled_dot_product_attention = scaled_dot_product_attention
diff --git a/rvc/ipex/gradscaler.py b/rvc/ipex/gradscaler.py
new file mode 100644
index 0000000000000000000000000000000000000000..a88fe225690c140435eeaa4cdb9aca01ad268518
--- /dev/null
+++ b/rvc/ipex/gradscaler.py
@@ -0,0 +1,188 @@
+from collections import defaultdict
+
+import torch
+import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+import intel_extension_for_pytorch._C as core # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+OptState = ipex.cpu.autocast._grad_scaler.OptState
+_MultiDeviceReplicator = ipex.cpu.autocast._grad_scaler._MultiDeviceReplicator
+_refresh_per_optimizer_state = (
+ ipex.cpu.autocast._grad_scaler._refresh_per_optimizer_state
+)
+
+
+def _unscale_grads_(
+ self, optimizer, inv_scale, found_inf, allow_fp16
+): # pylint: disable=unused-argument
+ per_device_inv_scale = _MultiDeviceReplicator(inv_scale)
+ per_device_found_inf = _MultiDeviceReplicator(found_inf)
+
+ # To set up _amp_foreach_non_finite_check_and_unscale_, split grads by device and dtype.
+ # There could be hundreds of grads, so we'd like to iterate through them just once.
+ # However, we don't know their devices or dtypes in advance.
+
+ # https://stackoverflow.com/questions/5029934/defaultdict-of-defaultdict
+ # Google says mypy struggles with defaultdicts type annotations.
+ per_device_and_dtype_grads = defaultdict(lambda: defaultdict(list)) # type: ignore[var-annotated]
+ # sync grad to master weight
+ if hasattr(optimizer, "sync_grad"):
+ optimizer.sync_grad()
+ with torch.no_grad():
+ for group in optimizer.param_groups:
+ for param in group["params"]:
+ if param.grad is None:
+ continue
+ if (not allow_fp16) and param.grad.dtype == torch.float16:
+ raise ValueError("Attempting to unscale FP16 gradients.")
+ if param.grad.is_sparse:
+ # is_coalesced() == False means the sparse grad has values with duplicate indices.
+ # coalesce() deduplicates indices and adds all values that have the same index.
+ # For scaled fp16 values, there's a good chance coalescing will cause overflow,
+ # so we should check the coalesced _values().
+ if param.grad.dtype is torch.float16:
+ param.grad = param.grad.coalesce()
+ to_unscale = param.grad._values()
+ else:
+ to_unscale = param.grad
+
+ # -: is there a way to split by device and dtype without appending in the inner loop?
+ to_unscale = to_unscale.to("cpu")
+ per_device_and_dtype_grads[to_unscale.device][to_unscale.dtype].append(
+ to_unscale
+ )
+
+ for _, per_dtype_grads in per_device_and_dtype_grads.items():
+ for grads in per_dtype_grads.values():
+ core._amp_foreach_non_finite_check_and_unscale_(
+ grads,
+ per_device_found_inf.get("cpu"),
+ per_device_inv_scale.get("cpu"),
+ )
+
+ return per_device_found_inf._per_device_tensors
+
+
+def unscale_(self, optimizer):
+ """
+ Divides ("unscales") the optimizer's gradient tensors by the scale factor.
+ :meth:`unscale_` is optional, serving cases where you need to
+ :ref:`modify or inspect gradients`
+ between the backward pass(es) and :meth:`step`.
+ If :meth:`unscale_` is not called explicitly, gradients will be unscaled automatically during :meth:`step`.
+ Simple example, using :meth:`unscale_` to enable clipping of unscaled gradients::
+ ...
+ scaler.scale(loss).backward()
+ scaler.unscale_(optimizer)
+ torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm)
+ scaler.step(optimizer)
+ scaler.update()
+ Args:
+ optimizer (torch.optim.Optimizer): Optimizer that owns the gradients to be unscaled.
+ .. warning::
+ :meth:`unscale_` should only be called once per optimizer per :meth:`step` call,
+ and only after all gradients for that optimizer's assigned parameters have been accumulated.
+ Calling :meth:`unscale_` twice for a given optimizer between each :meth:`step` triggers a RuntimeError.
+ .. warning::
+ :meth:`unscale_` may unscale sparse gradients out of place, replacing the ``.grad`` attribute.
+ """
+ if not self._enabled:
+ return
+
+ self._check_scale_growth_tracker("unscale_")
+
+ optimizer_state = self._per_optimizer_states[id(optimizer)]
+
+ if optimizer_state["stage"] is OptState.UNSCALED: # pylint: disable=no-else-raise
+ raise RuntimeError(
+ "unscale_() has already been called on this optimizer since the last update()."
+ )
+ elif optimizer_state["stage"] is OptState.STEPPED:
+ raise RuntimeError("unscale_() is being called after step().")
+
+ # FP32 division can be imprecise for certain compile options, so we carry out the reciprocal in FP64.
+ assert self._scale is not None
+ inv_scale = (
+ self._scale.to("cpu").double().reciprocal().float().to(self._scale.device)
+ )
+ found_inf = torch.full((1,), 0.0, dtype=torch.float32, device=self._scale.device)
+
+ optimizer_state["found_inf_per_device"] = self._unscale_grads_(
+ optimizer, inv_scale, found_inf, False
+ )
+ optimizer_state["stage"] = OptState.UNSCALED
+
+
+def update(self, new_scale=None):
+ """
+ Updates the scale factor.
+ If any optimizer steps were skipped the scale is multiplied by ``backoff_factor``
+ to reduce it. If ``growth_interval`` unskipped iterations occurred consecutively,
+ the scale is multiplied by ``growth_factor`` to increase it.
+ Passing ``new_scale`` sets the new scale value manually. (``new_scale`` is not
+ used directly, it's used to fill GradScaler's internal scale tensor. So if
+ ``new_scale`` was a tensor, later in-place changes to that tensor will not further
+ affect the scale GradScaler uses internally.)
+ Args:
+ new_scale (float or :class:`torch.FloatTensor`, optional, default=None): New scale factor.
+ .. warning::
+ :meth:`update` should only be called at the end of the iteration, after ``scaler.step(optimizer)`` has
+ been invoked for all optimizers used this iteration.
+ """
+ if not self._enabled:
+ return
+
+ _scale, _growth_tracker = self._check_scale_growth_tracker("update")
+
+ if new_scale is not None:
+ # Accept a new user-defined scale.
+ if isinstance(new_scale, float):
+ self._scale.fill_(new_scale) # type: ignore[union-attr]
+ else:
+ reason = "new_scale should be a float or a 1-element torch.FloatTensor with requires_grad=False."
+ assert isinstance(new_scale, torch.FloatTensor), reason # type: ignore[attr-defined]
+ assert new_scale.numel() == 1, reason
+ assert new_scale.requires_grad is False, reason
+ self._scale.copy_(new_scale) # type: ignore[union-attr]
+ else:
+ # Consume shared inf/nan data collected from optimizers to update the scale.
+ # If all found_inf tensors are on the same device as self._scale, this operation is asynchronous.
+ found_infs = [
+ found_inf.to(device="cpu", non_blocking=True)
+ for state in self._per_optimizer_states.values()
+ for found_inf in state["found_inf_per_device"].values()
+ ]
+
+ assert len(found_infs) > 0, "No inf checks were recorded prior to update."
+
+ found_inf_combined = found_infs[0]
+ if len(found_infs) > 1:
+ for i in range(1, len(found_infs)):
+ found_inf_combined += found_infs[i]
+
+ to_device = _scale.device
+ _scale = _scale.to("cpu")
+ _growth_tracker = _growth_tracker.to("cpu")
+
+ core._amp_update_scale_(
+ _scale,
+ _growth_tracker,
+ found_inf_combined,
+ self._growth_factor,
+ self._backoff_factor,
+ self._growth_interval,
+ )
+
+ _scale = _scale.to(to_device)
+ _growth_tracker = _growth_tracker.to(to_device)
+ # To prepare for next iteration, clear the data collected from optimizers this iteration.
+ self._per_optimizer_states = defaultdict(_refresh_per_optimizer_state)
+
+
+def gradscaler_init():
+ torch.xpu.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
+ torch.xpu.amp.GradScaler._unscale_grads_ = _unscale_grads_
+ torch.xpu.amp.GradScaler.unscale_ = unscale_
+ torch.xpu.amp.GradScaler.update = update
+ return torch.xpu.amp.GradScaler
diff --git a/rvc/ipex/hijacks.py b/rvc/ipex/hijacks.py
new file mode 100644
index 0000000000000000000000000000000000000000..51e72e7d8aae710eecb9c6461ef16bd543ce59f6
--- /dev/null
+++ b/rvc/ipex/hijacks.py
@@ -0,0 +1,366 @@
+import contextlib
+import importlib
+
+import torch
+import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long, unnecessary-lambda, no-else-return
+
+
+class CondFunc: # pylint: disable=missing-class-docstring
+ def __new__(cls, orig_func, sub_func, cond_func):
+ self = super(CondFunc, cls).__new__(cls)
+ if isinstance(orig_func, str):
+ func_path = orig_func.split(".")
+ for i in range(len(func_path) - 1, -1, -1):
+ try:
+ resolved_obj = importlib.import_module(".".join(func_path[:i]))
+ break
+ except ImportError:
+ pass
+ for attr_name in func_path[i:-1]:
+ resolved_obj = getattr(resolved_obj, attr_name)
+ orig_func = getattr(resolved_obj, func_path[-1])
+ setattr(
+ resolved_obj,
+ func_path[-1],
+ lambda *args, **kwargs: self(*args, **kwargs),
+ )
+ self.__init__(orig_func, sub_func, cond_func)
+ return lambda *args, **kwargs: self(*args, **kwargs)
+
+ def __init__(self, orig_func, sub_func, cond_func):
+ self.__orig_func = orig_func
+ self.__sub_func = sub_func
+ self.__cond_func = cond_func
+
+ def __call__(self, *args, **kwargs):
+ if not self.__cond_func or self.__cond_func(self.__orig_func, *args, **kwargs):
+ return self.__sub_func(self.__orig_func, *args, **kwargs)
+ else:
+ return self.__orig_func(*args, **kwargs)
+
+
+_utils = torch.utils.data._utils
+
+
+def _shutdown_workers(self):
+ if (
+ torch.utils.data._utils is None
+ or torch.utils.data._utils.python_exit_status is True
+ or torch.utils.data._utils.python_exit_status is None
+ ):
+ return
+ if hasattr(self, "_shutdown") and not self._shutdown:
+ self._shutdown = True
+ try:
+ if hasattr(self, "_pin_memory_thread"):
+ self._pin_memory_thread_done_event.set()
+ self._worker_result_queue.put((None, None))
+ self._pin_memory_thread.join()
+ self._worker_result_queue.cancel_join_thread()
+ self._worker_result_queue.close()
+ self._workers_done_event.set()
+ for worker_id in range(len(self._workers)):
+ if self._persistent_workers or self._workers_status[worker_id]:
+ self._mark_worker_as_unavailable(worker_id, shutdown=True)
+ for w in self._workers: # pylint: disable=invalid-name
+ w.join(timeout=torch.utils.data._utils.MP_STATUS_CHECK_INTERVAL)
+ for q in self._index_queues: # pylint: disable=invalid-name
+ q.cancel_join_thread()
+ q.close()
+ finally:
+ if self._worker_pids_set:
+ torch.utils.data._utils.signal_handling._remove_worker_pids(id(self))
+ self._worker_pids_set = False
+ for w in self._workers: # pylint: disable=invalid-name
+ if w.is_alive():
+ w.terminate()
+
+
+class DummyDataParallel(
+ torch.nn.Module
+): # pylint: disable=missing-class-docstring, unused-argument, too-few-public-methods
+ def __new__(
+ cls, module, device_ids=None, output_device=None, dim=0
+ ): # pylint: disable=unused-argument
+ if isinstance(device_ids, list) and len(device_ids) > 1:
+ print("IPEX backend doesn't support DataParallel on multiple XPU devices")
+ return module.to("xpu")
+
+
+def return_null_context(*args, **kwargs): # pylint: disable=unused-argument
+ return contextlib.nullcontext()
+
+
+def check_device(device):
+ return bool(
+ (isinstance(device, torch.device) and device.type == "cuda")
+ or (isinstance(device, str) and "cuda" in device)
+ or isinstance(device, int)
+ )
+
+
+def return_xpu(device):
+ return (
+ f"xpu:{device[-1]}"
+ if isinstance(device, str) and ":" in device
+ else (
+ f"xpu:{device}"
+ if isinstance(device, int)
+ else torch.device("xpu") if isinstance(device, torch.device) else "xpu"
+ )
+ )
+
+
+def ipex_no_cuda(orig_func, *args, **kwargs):
+ torch.cuda.is_available = lambda: False
+ orig_func(*args, **kwargs)
+ torch.cuda.is_available = torch.xpu.is_available
+
+
+original_autocast = torch.autocast
+
+
+def ipex_autocast(*args, **kwargs):
+ if len(args) > 0 and args[0] == "cuda":
+ return original_autocast("xpu", *args[1:], **kwargs)
+ else:
+ return original_autocast(*args, **kwargs)
+
+
+original_torch_cat = torch.cat
+
+
+def torch_cat(tensor, *args, **kwargs):
+ if len(tensor) == 3 and (
+ tensor[0].dtype != tensor[1].dtype or tensor[2].dtype != tensor[1].dtype
+ ):
+ return original_torch_cat(
+ [tensor[0].to(tensor[1].dtype), tensor[1], tensor[2].to(tensor[1].dtype)],
+ *args,
+ **kwargs,
+ )
+ else:
+ return original_torch_cat(tensor, *args, **kwargs)
+
+
+original_interpolate = torch.nn.functional.interpolate
+
+
+def interpolate(
+ tensor,
+ size=None,
+ scale_factor=None,
+ mode="nearest",
+ align_corners=None,
+ recompute_scale_factor=None,
+ antialias=False,
+): # pylint: disable=too-many-arguments
+ if antialias or align_corners is not None:
+ return_device = tensor.device
+ return_dtype = tensor.dtype
+ return original_interpolate(
+ tensor.to("cpu", dtype=torch.float32),
+ size=size,
+ scale_factor=scale_factor,
+ mode=mode,
+ align_corners=align_corners,
+ recompute_scale_factor=recompute_scale_factor,
+ antialias=antialias,
+ ).to(return_device, dtype=return_dtype)
+ else:
+ return original_interpolate(
+ tensor,
+ size=size,
+ scale_factor=scale_factor,
+ mode=mode,
+ align_corners=align_corners,
+ recompute_scale_factor=recompute_scale_factor,
+ antialias=antialias,
+ )
+
+
+original_linalg_solve = torch.linalg.solve
+
+
+def linalg_solve(A, B, *args, **kwargs): # pylint: disable=invalid-name
+ if A.device != torch.device("cpu") or B.device != torch.device("cpu"):
+ return_device = A.device
+ return original_linalg_solve(A.to("cpu"), B.to("cpu"), *args, **kwargs).to(
+ return_device
+ )
+ else:
+ return original_linalg_solve(A, B, *args, **kwargs)
+
+
+def ipex_hijacks():
+ CondFunc(
+ "torch.Tensor.to",
+ lambda orig_func, self, device=None, *args, **kwargs: orig_func(
+ self, return_xpu(device), *args, **kwargs
+ ),
+ lambda orig_func, self, device=None, *args, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.Tensor.cuda",
+ lambda orig_func, self, device=None, *args, **kwargs: orig_func(
+ self, return_xpu(device), *args, **kwargs
+ ),
+ lambda orig_func, self, device=None, *args, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.empty",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.load",
+ lambda orig_func, *args, map_location=None, **kwargs: orig_func(
+ *args, return_xpu(map_location), **kwargs
+ ),
+ lambda orig_func, *args, map_location=None, **kwargs: map_location is None
+ or check_device(map_location),
+ )
+ CondFunc(
+ "torch.randn",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.ones",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.zeros",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.tensor",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+ CondFunc(
+ "torch.linspace",
+ lambda orig_func, *args, device=None, **kwargs: orig_func(
+ *args, device=return_xpu(device), **kwargs
+ ),
+ lambda orig_func, *args, device=None, **kwargs: check_device(device),
+ )
+
+ CondFunc(
+ "torch.Generator",
+ lambda orig_func, device=None: torch.xpu.Generator(device),
+ lambda orig_func, device=None: device is not None
+ and device != torch.device("cpu")
+ and device != "cpu",
+ )
+
+ CondFunc(
+ "torch.batch_norm",
+ lambda orig_func, input, weight, bias, *args, **kwargs: orig_func(
+ input,
+ (
+ weight
+ if weight is not None
+ else torch.ones(input.size()[1], device=input.device)
+ ),
+ (
+ bias
+ if bias is not None
+ else torch.zeros(input.size()[1], device=input.device)
+ ),
+ *args,
+ **kwargs,
+ ),
+ lambda orig_func, input, *args, **kwargs: input.device != torch.device("cpu"),
+ )
+ CondFunc(
+ "torch.instance_norm",
+ lambda orig_func, input, weight, bias, *args, **kwargs: orig_func(
+ input,
+ (
+ weight
+ if weight is not None
+ else torch.ones(input.size()[1], device=input.device)
+ ),
+ (
+ bias
+ if bias is not None
+ else torch.zeros(input.size()[1], device=input.device)
+ ),
+ *args,
+ **kwargs,
+ ),
+ lambda orig_func, input, *args, **kwargs: input.device != torch.device("cpu"),
+ )
+
+ # Functions with dtype errors:
+ CondFunc(
+ "torch.nn.modules.GroupNorm.forward",
+ lambda orig_func, self, input: orig_func(
+ self, input.to(self.weight.data.dtype)
+ ),
+ lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+ )
+ CondFunc(
+ "torch.nn.modules.linear.Linear.forward",
+ lambda orig_func, self, input: orig_func(
+ self, input.to(self.weight.data.dtype)
+ ),
+ lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+ )
+ CondFunc(
+ "torch.nn.modules.conv.Conv2d.forward",
+ lambda orig_func, self, input: orig_func(
+ self, input.to(self.weight.data.dtype)
+ ),
+ lambda orig_func, self, input: input.dtype != self.weight.data.dtype,
+ )
+ CondFunc(
+ "torch.nn.functional.layer_norm",
+ lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs: orig_func(
+ input.to(weight.data.dtype), normalized_shape, weight, *args, **kwargs
+ ),
+ lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs: weight
+ is not None
+ and input.dtype != weight.data.dtype,
+ )
+
+ # Diffusers Float64 (ARC GPUs doesn't support double or Float64):
+ if not torch.xpu.has_fp64_dtype():
+ CondFunc(
+ "torch.from_numpy",
+ lambda orig_func, ndarray: orig_func(ndarray.astype("float32")),
+ lambda orig_func, ndarray: ndarray.dtype == float,
+ )
+
+ # Broken functions when torch.cuda.is_available is True:
+ CondFunc(
+ "torch.utils.data.dataloader._BaseDataLoaderIter.__init__",
+ lambda orig_func, *args, **kwargs: ipex_no_cuda(orig_func, *args, **kwargs),
+ lambda orig_func, *args, **kwargs: True,
+ )
+
+ # Functions that make compile mad with CondFunc:
+ torch.utils.data.dataloader._MultiProcessingDataLoaderIter._shutdown_workers = (
+ _shutdown_workers
+ )
+ torch.nn.DataParallel = DummyDataParallel
+ torch.autocast = ipex_autocast
+ torch.cat = torch_cat
+ torch.linalg.solve = linalg_solve
+ torch.nn.functional.interpolate = interpolate
+ torch.backends.cuda.sdp_kernel = return_null_context
diff --git a/rvc/ipex/init.py b/rvc/ipex/init.py
new file mode 100644
index 0000000000000000000000000000000000000000..82b93f63ab725a62f3ea3b1c840f4f6e1ad53fdf
--- /dev/null
+++ b/rvc/ipex/init.py
@@ -0,0 +1,192 @@
+import os
+import sys
+import contextlib
+
+import torch
+import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+
+from .hijacks import ipex_hijacks
+from .attention import attention_init
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+
+def ipex_init(): # pylint: disable=too-many-statements
+ try:
+ # Replace cuda with xpu:
+ torch.cuda.current_device = torch.xpu.current_device
+ torch.cuda.current_stream = torch.xpu.current_stream
+ torch.cuda.device = torch.xpu.device
+ torch.cuda.device_count = torch.xpu.device_count
+ torch.cuda.device_of = torch.xpu.device_of
+ torch.cuda.get_device_name = torch.xpu.get_device_name
+ torch.cuda.get_device_properties = torch.xpu.get_device_properties
+ torch.cuda.init = torch.xpu.init
+ torch.cuda.is_available = torch.xpu.is_available
+ torch.cuda.is_initialized = torch.xpu.is_initialized
+ torch.cuda.is_current_stream_capturing = lambda: False
+ torch.cuda.set_device = torch.xpu.set_device
+ torch.cuda.stream = torch.xpu.stream
+ torch.cuda.synchronize = torch.xpu.synchronize
+ torch.cuda.Event = torch.xpu.Event
+ torch.cuda.Stream = torch.xpu.Stream
+ torch.cuda.FloatTensor = torch.xpu.FloatTensor
+ torch.Tensor.cuda = torch.Tensor.xpu
+ torch.Tensor.is_cuda = torch.Tensor.is_xpu
+ torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
+ torch.cuda._initialized = torch.xpu.lazy_init._initialized
+ torch.cuda._lazy_seed_tracker = torch.xpu.lazy_init._lazy_seed_tracker
+ torch.cuda._queued_calls = torch.xpu.lazy_init._queued_calls
+ torch.cuda._tls = torch.xpu.lazy_init._tls
+ torch.cuda.threading = torch.xpu.lazy_init.threading
+ torch.cuda.traceback = torch.xpu.lazy_init.traceback
+ torch.cuda.Optional = torch.xpu.Optional
+ torch.cuda.__cached__ = torch.xpu.__cached__
+ torch.cuda.__loader__ = torch.xpu.__loader__
+ torch.cuda.ComplexFloatStorage = torch.xpu.ComplexFloatStorage
+ torch.cuda.Tuple = torch.xpu.Tuple
+ torch.cuda.streams = torch.xpu.streams
+ torch.cuda._lazy_new = torch.xpu._lazy_new
+ torch.cuda.FloatStorage = torch.xpu.FloatStorage
+ torch.cuda.Any = torch.xpu.Any
+ torch.cuda.__doc__ = torch.xpu.__doc__
+ torch.cuda.default_generators = torch.xpu.default_generators
+ torch.cuda.HalfTensor = torch.xpu.HalfTensor
+ torch.cuda._get_device_index = torch.xpu._get_device_index
+ torch.cuda.__path__ = torch.xpu.__path__
+ torch.cuda.Device = torch.xpu.Device
+ torch.cuda.IntTensor = torch.xpu.IntTensor
+ torch.cuda.ByteStorage = torch.xpu.ByteStorage
+ torch.cuda.set_stream = torch.xpu.set_stream
+ torch.cuda.BoolStorage = torch.xpu.BoolStorage
+ torch.cuda.os = torch.xpu.os
+ torch.cuda.torch = torch.xpu.torch
+ torch.cuda.BFloat16Storage = torch.xpu.BFloat16Storage
+ torch.cuda.Union = torch.xpu.Union
+ torch.cuda.DoubleTensor = torch.xpu.DoubleTensor
+ torch.cuda.ShortTensor = torch.xpu.ShortTensor
+ torch.cuda.LongTensor = torch.xpu.LongTensor
+ torch.cuda.IntStorage = torch.xpu.IntStorage
+ torch.cuda.LongStorage = torch.xpu.LongStorage
+ torch.cuda.__annotations__ = torch.xpu.__annotations__
+ torch.cuda.__package__ = torch.xpu.__package__
+ torch.cuda.__builtins__ = torch.xpu.__builtins__
+ torch.cuda.CharTensor = torch.xpu.CharTensor
+ torch.cuda.List = torch.xpu.List
+ torch.cuda._lazy_init = torch.xpu._lazy_init
+ torch.cuda.BFloat16Tensor = torch.xpu.BFloat16Tensor
+ torch.cuda.DoubleStorage = torch.xpu.DoubleStorage
+ torch.cuda.ByteTensor = torch.xpu.ByteTensor
+ torch.cuda.StreamContext = torch.xpu.StreamContext
+ torch.cuda.ComplexDoubleStorage = torch.xpu.ComplexDoubleStorage
+ torch.cuda.ShortStorage = torch.xpu.ShortStorage
+ torch.cuda._lazy_call = torch.xpu._lazy_call
+ torch.cuda.HalfStorage = torch.xpu.HalfStorage
+ torch.cuda.random = torch.xpu.random
+ torch.cuda._device = torch.xpu._device
+ torch.cuda.classproperty = torch.xpu.classproperty
+ torch.cuda.__name__ = torch.xpu.__name__
+ torch.cuda._device_t = torch.xpu._device_t
+ torch.cuda.warnings = torch.xpu.warnings
+ torch.cuda.__spec__ = torch.xpu.__spec__
+ torch.cuda.BoolTensor = torch.xpu.BoolTensor
+ torch.cuda.CharStorage = torch.xpu.CharStorage
+ torch.cuda.__file__ = torch.xpu.__file__
+ torch.cuda._is_in_bad_fork = torch.xpu.lazy_init._is_in_bad_fork
+ # torch.cuda.is_current_stream_capturing = torch.xpu.is_current_stream_capturing
+
+ # Memory:
+ torch.cuda.memory = torch.xpu.memory
+ if "linux" in sys.platform and "WSL2" in os.popen("uname -a").read():
+ torch.xpu.empty_cache = lambda: None
+ torch.cuda.empty_cache = torch.xpu.empty_cache
+ torch.cuda.memory_stats = torch.xpu.memory_stats
+ torch.cuda.memory_summary = torch.xpu.memory_summary
+ torch.cuda.memory_snapshot = torch.xpu.memory_snapshot
+ torch.cuda.memory_allocated = torch.xpu.memory_allocated
+ torch.cuda.max_memory_allocated = torch.xpu.max_memory_allocated
+ torch.cuda.memory_reserved = torch.xpu.memory_reserved
+ torch.cuda.memory_cached = torch.xpu.memory_reserved
+ torch.cuda.max_memory_reserved = torch.xpu.max_memory_reserved
+ torch.cuda.max_memory_cached = torch.xpu.max_memory_reserved
+ torch.cuda.reset_peak_memory_stats = torch.xpu.reset_peak_memory_stats
+ torch.cuda.reset_max_memory_cached = torch.xpu.reset_peak_memory_stats
+ torch.cuda.reset_max_memory_allocated = torch.xpu.reset_peak_memory_stats
+ torch.cuda.memory_stats_as_nested_dict = torch.xpu.memory_stats_as_nested_dict
+ torch.cuda.reset_accumulated_memory_stats = (
+ torch.xpu.reset_accumulated_memory_stats
+ )
+
+ # RNG:
+ torch.cuda.get_rng_state = torch.xpu.get_rng_state
+ torch.cuda.get_rng_state_all = torch.xpu.get_rng_state_all
+ torch.cuda.set_rng_state = torch.xpu.set_rng_state
+ torch.cuda.set_rng_state_all = torch.xpu.set_rng_state_all
+ torch.cuda.manual_seed = torch.xpu.manual_seed
+ torch.cuda.manual_seed_all = torch.xpu.manual_seed_all
+ torch.cuda.seed = torch.xpu.seed
+ torch.cuda.seed_all = torch.xpu.seed_all
+ torch.cuda.initial_seed = torch.xpu.initial_seed
+
+ # AMP:
+ torch.cuda.amp = torch.xpu.amp
+ if not hasattr(torch.cuda.amp, "common"):
+ torch.cuda.amp.common = contextlib.nullcontext()
+ torch.cuda.amp.common.amp_definitely_not_available = lambda: False
+ try:
+ torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
+ except Exception: # pylint: disable=broad-exception-caught
+ try:
+ from .gradscaler import (
+ gradscaler_init,
+ ) # pylint: disable=import-outside-toplevel, import-error
+
+ gradscaler_init()
+ torch.cuda.amp.GradScaler = torch.xpu.amp.GradScaler
+ except Exception: # pylint: disable=broad-exception-caught
+ torch.cuda.amp.GradScaler = ipex.cpu.autocast._grad_scaler.GradScaler
+
+ # C
+ torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentStream
+ ipex._C._DeviceProperties.major = 2023
+ ipex._C._DeviceProperties.minor = 2
+
+ # Fix functions with ipex:
+ torch.cuda.mem_get_info = lambda device=None: [
+ (
+ torch.xpu.get_device_properties(device).total_memory
+ - torch.xpu.memory_allocated(device)
+ ),
+ torch.xpu.get_device_properties(device).total_memory,
+ ]
+ torch._utils._get_available_device_type = lambda: "xpu"
+ torch.has_cuda = True
+ torch.cuda.has_half = True
+ torch.cuda.is_bf16_supported = lambda *args, **kwargs: True
+ torch.cuda.is_fp16_supported = lambda *args, **kwargs: True
+ torch.version.cuda = "11.7"
+ torch.cuda.get_device_capability = lambda *args, **kwargs: [11, 7]
+ torch.cuda.get_device_properties.major = 11
+ torch.cuda.get_device_properties.minor = 7
+ torch.cuda.ipc_collect = lambda *args, **kwargs: None
+ torch.cuda.utilization = lambda *args, **kwargs: 0
+ if hasattr(torch.xpu, "getDeviceIdListForCard"):
+ torch.cuda.getDeviceIdListForCard = torch.xpu.getDeviceIdListForCard
+ torch.cuda.get_device_id_list_per_card = torch.xpu.getDeviceIdListForCard
+ else:
+ torch.cuda.getDeviceIdListForCard = torch.xpu.get_device_id_list_per_card
+ torch.cuda.get_device_id_list_per_card = (
+ torch.xpu.get_device_id_list_per_card
+ )
+
+ ipex_hijacks()
+ attention_init()
+ try:
+ from .diffusers import ipex_diffusers
+
+ ipex_diffusers()
+ except Exception: # pylint: disable=broad-exception-caught
+ pass
+ except Exception as e:
+ return False, e
+ return True, None
diff --git a/rvc/jit/__init__.py b/rvc/jit/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..9016b939d8b131195b4be54b3605f9e60db7e13d
--- /dev/null
+++ b/rvc/jit/__init__.py
@@ -0,0 +1 @@
+from .jit import load_inputs, get_jit_model, export_jit_model, save_pickle
diff --git a/rvc/jit/jit.py b/rvc/jit/jit.py
new file mode 100644
index 0000000000000000000000000000000000000000..1da0f2cdb76f0eac41476b9aa63c4f767023ad29
--- /dev/null
+++ b/rvc/jit/jit.py
@@ -0,0 +1,78 @@
+import pickle
+from io import BytesIO
+from collections import OrderedDict
+import os
+
+import torch
+
+
+def load_pickle(path: str):
+ with open(path, "rb") as f:
+ return pickle.load(f)
+
+
+def save_pickle(ckpt: dict, save_path: str):
+ with open(save_path, "wb") as f:
+ pickle.dump(ckpt, f)
+
+
+def load_inputs(path: torch.serialization.FILE_LIKE, device: str, is_half=False):
+ parm = torch.load(path, map_location=torch.device("cpu"))
+ for key in parm.keys():
+ parm[key] = parm[key].to(device)
+ if is_half and parm[key].dtype == torch.float32:
+ parm[key] = parm[key].half()
+ elif not is_half and parm[key].dtype == torch.float16:
+ parm[key] = parm[key].float()
+ return parm
+
+
+def export_jit_model(
+ model: torch.nn.Module,
+ mode: str = "trace",
+ inputs: dict = None,
+ device=torch.device("cpu"),
+ is_half: bool = False,
+) -> dict:
+ model = model.half() if is_half else model.float()
+ model.eval()
+ if mode == "trace":
+ assert inputs is not None
+ model_jit = torch.jit.trace(model, example_kwarg_inputs=inputs)
+ elif mode == "script":
+ model_jit = torch.jit.script(model)
+ model_jit.to(device)
+ model_jit = model_jit.half() if is_half else model_jit.float()
+ buffer = BytesIO()
+ # model_jit=model_jit.cpu()
+ torch.jit.save(model_jit, buffer)
+ del model_jit
+ cpt = OrderedDict()
+ cpt["model"] = buffer.getvalue()
+ cpt["is_half"] = is_half
+ return cpt
+
+
+def get_jit_model(model_path: str, is_half: bool, device: str, exporter):
+ jit_model_path = model_path.rstrip(".pth")
+ jit_model_path += ".half.jit" if is_half else ".jit"
+ ckpt = None
+
+ if os.path.exists(jit_model_path):
+ ckpt = load_pickle(jit_model_path)
+ model_device = ckpt["device"]
+ if model_device != str(device):
+ del ckpt
+ ckpt = None
+
+ if ckpt is None:
+ ckpt = exporter(
+ model_path=model_path,
+ mode="script",
+ inputs_path=None,
+ save_path=jit_model_path,
+ device=device,
+ is_half=is_half,
+ )
+
+ return ckpt
diff --git a/rvc/layers/__init__.py b/rvc/layers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/rvc/layers/attentions.py b/rvc/layers/attentions.py
new file mode 100644
index 0000000000000000000000000000000000000000..22b626dad7f0dcd4b09f7cf4a85c647940ab9ffc
--- /dev/null
+++ b/rvc/layers/attentions.py
@@ -0,0 +1,292 @@
+import math
+from typing import Optional
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+
+class MultiHeadAttention(nn.Module):
+ def __init__(
+ self,
+ channels: int,
+ out_channels: int,
+ n_heads: int,
+ p_dropout: float = 0.0,
+ window_size: Optional[int] = None,
+ heads_share: bool = True,
+ block_length: Optional[int] = None,
+ proximal_bias: bool = False,
+ proximal_init: bool = False,
+ ):
+ super(MultiHeadAttention, self).__init__()
+ assert channels % n_heads == 0
+
+ self.channels = channels
+ self.out_channels = out_channels
+ self.n_heads = n_heads
+ self.p_dropout = p_dropout
+ self.window_size = window_size
+ self.heads_share = heads_share
+ self.block_length = block_length
+ self.proximal_bias = proximal_bias
+ self.proximal_init = proximal_init
+ self.attn = None
+
+ self.k_channels = channels // n_heads
+ self.conv_q = nn.Conv1d(channels, channels, 1)
+ self.conv_k = nn.Conv1d(channels, channels, 1)
+ self.conv_v = nn.Conv1d(channels, channels, 1)
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
+ self.drop = nn.Dropout(p_dropout)
+
+ if window_size is not None:
+ n_heads_rel = 1 if heads_share else n_heads
+ rel_stddev = self.k_channels**-0.5
+ self.emb_rel_k = nn.Parameter(
+ torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+ * rel_stddev
+ )
+ self.emb_rel_v = nn.Parameter(
+ torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
+ * rel_stddev
+ )
+
+ nn.init.xavier_uniform_(self.conv_q.weight)
+ nn.init.xavier_uniform_(self.conv_k.weight)
+ nn.init.xavier_uniform_(self.conv_v.weight)
+ if proximal_init:
+ with torch.no_grad():
+ self.conv_k.weight.copy_(self.conv_q.weight)
+ self.conv_k.bias.copy_(self.conv_q.bias)
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ c: torch.Tensor,
+ attn_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, c, attn_mask=attn_mask)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ c: torch.Tensor,
+ attn_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ q = self.conv_q(x)
+ k = self.conv_k(c)
+ v = self.conv_v(c)
+
+ x, _ = self._attention(q, k, v, mask=attn_mask)
+
+ x = self.conv_o(x)
+ return x
+
+ def _attention(
+ self,
+ query: torch.Tensor,
+ key: torch.Tensor,
+ value: torch.Tensor,
+ mask: Optional[torch.Tensor] = None,
+ ):
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
+ b, d, t_s = key.size()
+ t_t = query.size(2)
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
+
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
+ if self.window_size is not None:
+ assert (
+ t_s == t_t
+ ), "Relative attention is only available for self-attention."
+ key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+ rel_logits = self._matmul_with_relative_keys(
+ query / math.sqrt(self.k_channels), key_relative_embeddings
+ )
+ scores_local = self._relative_position_to_absolute_position(rel_logits)
+ scores = scores + scores_local
+ if self.proximal_bias:
+ assert t_s == t_t, "Proximal bias is only available for self-attention."
+ scores = scores + self._attention_bias_proximal(t_s).to(
+ device=scores.device, dtype=scores.dtype
+ )
+ if mask is not None:
+ scores = scores.masked_fill(mask == 0, -1e4)
+ if self.block_length is not None:
+ assert (
+ t_s == t_t
+ ), "Local attention is only available for self-attention."
+ block_mask = (
+ torch.ones_like(scores)
+ .triu(-self.block_length)
+ .tril(self.block_length)
+ )
+ scores = scores.masked_fill(block_mask == 0, -1e4)
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
+ p_attn = self.drop(p_attn)
+ output = torch.matmul(p_attn, value)
+ if self.window_size is not None:
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
+ value_relative_embeddings = self._get_relative_embeddings(
+ self.emb_rel_v, t_s
+ )
+ output = output + self._matmul_with_relative_values(
+ relative_weights, value_relative_embeddings
+ )
+ output = (
+ output.transpose(2, 3).contiguous().view(b, d, t_t)
+ ) # [b, n_h, t_t, d_k] -> [b, d, t_t]
+ return output, p_attn
+
+ def _matmul_with_relative_values(self, x, y):
+ """
+ x: [b, h, l, m]
+ y: [h or 1, m, d]
+ ret: [b, h, l, d]
+ """
+ ret = torch.matmul(x, y.unsqueeze(0))
+ return ret
+
+ def _matmul_with_relative_keys(self, x, y):
+ """
+ x: [b, h, l, d]
+ y: [h or 1, m, d]
+ ret: [b, h, l, m]
+ """
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
+ return ret
+
+ def _get_relative_embeddings(self, relative_embeddings, length: int):
+ # max_relative_position = 2 * self.window_size + 1
+ # Pad first before slice to avoid using cond ops.
+ pad_length: int = max(length - (self.window_size + 1), 0)
+ slice_start_position = max((self.window_size + 1) - length, 0)
+ slice_end_position = slice_start_position + 2 * length - 1
+ if pad_length > 0:
+ padded_relative_embeddings = F.pad(
+ relative_embeddings,
+ [0, 0, pad_length, pad_length, 0, 0],
+ )
+ else:
+ padded_relative_embeddings = relative_embeddings
+ used_relative_embeddings = padded_relative_embeddings[
+ :, slice_start_position:slice_end_position
+ ]
+ return used_relative_embeddings
+
+ def _relative_position_to_absolute_position(self, x):
+ """
+ x: [b, h, l, 2*l-1]
+ ret: [b, h, l, l]
+ """
+ batch, heads, length, _ = x.size()
+ # Concat columns of pad to shift from relative to absolute indexing.
+ x = F.pad(
+ x,
+ [0, 1, 0, 0, 0, 0, 0, 0],
+ )
+
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
+ x_flat = x.view([batch, heads, length * 2 * length])
+ x_flat = F.pad(x_flat, [0, length - 1, 0, 0, 0, 0])
+
+ # Reshape and slice out the padded elements.
+ x_final = x_flat.view([batch, heads, length + 1, 2 * length - 1])[
+ :, :, :length, length - 1 :
+ ]
+ return x_final
+
+ def _absolute_position_to_relative_position(self, x):
+ """
+ x: [b, h, l, l]
+ ret: [b, h, l, 2*l-1]
+ """
+ batch, heads, length, _ = x.size()
+ # padd along column
+ x = F.pad(x, [0, length - 1, 0, 0, 0, 0, 0, 0])
+ x_flat = x.view([batch, heads, (length**2) + (length * (length - 1))])
+ # add 0's in the beginning that will skew the elements after reshape
+ x_flat = F.pad(x_flat, [length, 0, 0, 0, 0, 0])
+ x_final = x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
+ return x_final
+
+ def _attention_bias_proximal(self, length: int):
+ """Bias for self-attention to encourage attention to close positions.
+ Args:
+ length: an integer scalar.
+ Returns:
+ a Tensor with shape [1, 1, length, length]
+ """
+ r = torch.arange(length, dtype=torch.float32)
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
+
+
+class FFN(nn.Module):
+ """
+ Feed-Forward Network
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ filter_channels: int,
+ kernel_size: int,
+ p_dropout: float = 0.0,
+ activation: Optional[str] = None,
+ causal: bool = False,
+ ):
+ super(FFN, self).__init__()
+ self.in_channels = in_channels
+ self.out_channels = out_channels
+ self.filter_channels = filter_channels
+ self.kernel_size = kernel_size
+ self.p_dropout = p_dropout
+ self.activation = activation
+ self.causal = causal
+ self.is_activation = True if activation == "gelu" else False
+
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
+ self.drop = nn.Dropout(p_dropout)
+
+ def __call__(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+ return super().__call__(x, x_mask)
+
+ def forward(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+ x = self.conv_1(self._padding(x, x_mask))
+ if self.is_activation:
+ x = x * torch.sigmoid(1.702 * x)
+ else:
+ x = torch.relu(x)
+ x = self.drop(x)
+
+ x = self.conv_2(self._padding(x, x_mask))
+ return x * x_mask
+
+ def _padding(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+ if self.causal:
+ return self._causal_padding(x * x_mask)
+ return self._same_padding(x * x_mask)
+
+ def _causal_padding(self, x):
+ if self.kernel_size == 1:
+ return x
+ pad_l: int = self.kernel_size - 1
+ pad_r: int = 0
+ # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+ x = F.pad(x, [pad_l, pad_r, 0, 0, 0, 0])
+ return x
+
+ def _same_padding(self, x):
+ if self.kernel_size == 1:
+ return x
+ pad_l: int = (self.kernel_size - 1) // 2
+ pad_r: int = self.kernel_size // 2
+ # padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+ x = F.pad(x, [pad_l, pad_r, 0, 0, 0, 0])
+ return x
diff --git a/rvc/layers/discriminators.py b/rvc/layers/discriminators.py
new file mode 100644
index 0000000000000000000000000000000000000000..29e2dffed619fe3bc2e48ad97a5c03b410b60643
--- /dev/null
+++ b/rvc/layers/discriminators.py
@@ -0,0 +1,172 @@
+from typing import List, Tuple
+
+import torch
+from torch import nn
+from torch.nn import Conv1d, Conv2d
+from torch.nn import functional as F
+from torch.nn.utils import spectral_norm, weight_norm
+
+from .residuals import LRELU_SLOPE
+from .utils import get_padding
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+ """
+ version: 'v1' or 'v2'
+ """
+
+ def __init__(
+ self, version: str, use_spectral_norm: bool = False, has_xpu: bool = False
+ ):
+ super(MultiPeriodDiscriminator, self).__init__()
+ periods = (
+ (2, 3, 5, 7, 11, 17) if version == "v1" else (2, 3, 5, 7, 11, 17, 23, 37)
+ )
+
+ self.discriminators = nn.ModuleList(
+ [
+ DiscriminatorS(use_spectral_norm=use_spectral_norm),
+ *(
+ DiscriminatorP(
+ i, use_spectral_norm=use_spectral_norm, has_xpu=has_xpu
+ )
+ for i in periods
+ ),
+ ]
+ )
+
+ def __call__(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[
+ List[torch.Tensor],
+ List[torch.Tensor],
+ List[List[torch.Tensor]],
+ List[List[torch.Tensor]],
+ ]:
+ return super().__call__(y, y_hat)
+
+ def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[
+ List[torch.Tensor],
+ List[torch.Tensor],
+ List[List[torch.Tensor]],
+ List[List[torch.Tensor]],
+ ]:
+ y_d_rs = []
+ y_d_gs = []
+ fmap_rs = []
+ fmap_gs = []
+
+ for d in self.discriminators:
+ y_d_r, fmap_r = d(y)
+ y_d_g, fmap_g = d(y_hat)
+ y_d_rs.append(y_d_r)
+ y_d_gs.append(y_d_g)
+ fmap_rs.append(fmap_r)
+ fmap_gs.append(fmap_g)
+
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorS(torch.nn.Module):
+ def __init__(self, use_spectral_norm: bool = False):
+ super(DiscriminatorS, self).__init__()
+ norm_f = spectral_norm if use_spectral_norm else weight_norm
+
+ self.convs = nn.ModuleList(
+ [
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
+ ]
+ )
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
+
+ def __call__(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ return super().__call__(x)
+
+ def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ fmap = []
+
+ for l in self.convs:
+ x = l(x)
+ x = F.leaky_relu(x, LRELU_SLOPE)
+ fmap.append(x)
+
+ x = self.conv_post(x)
+ fmap.append(x)
+ x = torch.flatten(x, 1, -1)
+
+ return x, fmap
+
+
+class DiscriminatorP(torch.nn.Module):
+ def __init__(
+ self,
+ period: int,
+ kernel_size: int = 5,
+ stride: int = 3,
+ use_spectral_norm: bool = False,
+ has_xpu: bool = False,
+ ):
+ super(DiscriminatorP, self).__init__()
+ self.period = period
+ self.has_xpu = has_xpu
+ norm_f = spectral_norm if use_spectral_norm else weight_norm
+ sequence = (1, 32, 128, 512, 1024)
+ convs_padding = (get_padding(kernel_size, 1), 0)
+
+ self.convs = nn.ModuleList()
+ for i in range(len(sequence) - 1):
+ self.convs.append(
+ norm_f(
+ Conv2d(
+ sequence[i],
+ sequence[i + 1],
+ (kernel_size, 1),
+ (stride, 1),
+ padding=convs_padding,
+ )
+ )
+ )
+ self.convs.append(
+ norm_f(
+ Conv2d(
+ 1024,
+ 1024,
+ (kernel_size, 1),
+ 1,
+ padding=convs_padding,
+ )
+ )
+ )
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+ def __call__(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ return super().__call__(x)
+
+ def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+ fmap = []
+
+ # 1d to 2d
+ b, c, t = x.shape
+ if t % self.period != 0: # pad first
+ n_pad = self.period - (t % self.period)
+ if self.has_xpu and x.dtype == torch.bfloat16:
+ x = F.pad(x.to(dtype=torch.float16), (0, n_pad), "reflect").to(
+ dtype=torch.bfloat16
+ )
+ else:
+ x = F.pad(x, (0, n_pad), "reflect")
+ t = t + n_pad
+ x = x.view(b, c, t // self.period, self.period)
+
+ for l in self.convs:
+ x = l(x)
+ x = F.leaky_relu(x, LRELU_SLOPE)
+ fmap.append(x)
+ x = self.conv_post(x)
+ fmap.append(x)
+ x = torch.flatten(x, 1, -1)
+
+ return x, fmap
diff --git a/rvc/layers/encoders.py b/rvc/layers/encoders.py
new file mode 100644
index 0000000000000000000000000000000000000000..1fc2478b4887760ef78dc87b1491d4e4bd68b2cf
--- /dev/null
+++ b/rvc/layers/encoders.py
@@ -0,0 +1,221 @@
+import math
+from typing import Tuple, Optional
+
+import torch
+from torch import nn
+
+from .attentions import MultiHeadAttention, FFN
+from .norms import LayerNorm, WN
+from .utils import sequence_mask
+
+
+class Encoder(nn.Module):
+ def __init__(
+ self,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int = 1,
+ p_dropout: float = 0.0,
+ window_size: int = 10,
+ ):
+ super(Encoder, self).__init__()
+
+ self.hidden_channels = hidden_channels
+ self.filter_channels = filter_channels
+ self.n_heads = n_heads
+ self.n_layers = n_layers
+ self.kernel_size = kernel_size
+ self.p_dropout = p_dropout
+ self.window_size = window_size
+
+ self.drop = nn.Dropout(p_dropout)
+ self.attn_layers = nn.ModuleList()
+ self.norm_layers_1 = nn.ModuleList()
+ self.ffn_layers = nn.ModuleList()
+ self.norm_layers_2 = nn.ModuleList()
+
+ for _ in range(self.n_layers):
+ self.attn_layers.append(
+ MultiHeadAttention(
+ hidden_channels,
+ hidden_channels,
+ n_heads,
+ p_dropout=p_dropout,
+ window_size=window_size,
+ )
+ )
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
+ self.ffn_layers.append(
+ FFN(
+ hidden_channels,
+ hidden_channels,
+ filter_channels,
+ kernel_size,
+ p_dropout=p_dropout,
+ )
+ )
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+ def __call__(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+ return super().__call__(x, x_mask)
+
+ def forward(self, x: torch.Tensor, x_mask: torch.Tensor) -> torch.Tensor:
+ attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+ x = x * x_mask
+ for attn, norm1, ffn, norm2 in zip(
+ self.attn_layers,
+ self.norm_layers_1,
+ self.ffn_layers,
+ self.norm_layers_2,
+ ):
+ y = attn(x, x, attn_mask)
+ y = self.drop(y)
+ x = norm1(x + y)
+
+ y = ffn(x, x_mask)
+ y = self.drop(y)
+ x = norm2(x + y)
+ x = x * x_mask
+ return x
+
+
+class TextEncoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: float,
+ f0: bool = True,
+ ):
+ super(TextEncoder, self).__init__()
+
+ self.out_channels = out_channels
+ self.hidden_channels = hidden_channels
+ self.filter_channels = filter_channels
+ self.n_heads = n_heads
+ self.n_layers = n_layers
+ self.kernel_size = kernel_size
+ self.p_dropout = float(p_dropout)
+
+ self.emb_phone = nn.Linear(in_channels, hidden_channels)
+ self.lrelu = nn.LeakyReLU(0.1, inplace=True)
+ if f0 == True:
+ self.emb_pitch = nn.Embedding(256, hidden_channels) # pitch 256
+ self.encoder = Encoder(
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ float(p_dropout),
+ )
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+ def __call__(
+ self,
+ phone: torch.Tensor,
+ pitch: torch.Tensor,
+ lengths: torch.Tensor,
+ skip_head: Optional[int] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+ return super().__call__(
+ phone,
+ pitch,
+ lengths,
+ skip_head=skip_head,
+ )
+
+ def forward(
+ self,
+ phone: torch.Tensor,
+ pitch: torch.Tensor,
+ lengths: torch.Tensor,
+ skip_head: Optional[int] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+ x = self.emb_phone(phone)
+ if pitch is not None:
+ x += self.emb_pitch(pitch)
+ x = x * math.sqrt(self.hidden_channels) # [b, t, h]
+ x = self.lrelu(x)
+ x = torch.transpose(x, 1, -1) # [b, h, t]
+ x_mask = torch.unsqueeze(
+ sequence_mask(lengths, x.size(2)),
+ 1,
+ ).to(x.dtype)
+ x = self.encoder(x * x_mask, x_mask)
+ if skip_head is not None:
+ head = int(skip_head)
+ x = x[:, :, head:]
+ x_mask = x_mask[:, :, head:]
+ stats: torch.Tensor = self.proj(x) * x_mask
+ m, logs = torch.split(stats, self.out_channels, dim=1)
+ return m, logs, x_mask
+
+
+class PosteriorEncoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ hidden_channels: int,
+ kernel_size: int,
+ dilation_rate: int,
+ n_layers: int,
+ gin_channels=0,
+ ):
+ super(PosteriorEncoder, self).__init__()
+ self.in_channels = in_channels
+ self.out_channels = out_channels
+ self.hidden_channels = hidden_channels
+ self.kernel_size = kernel_size
+ self.dilation_rate = dilation_rate
+ self.n_layers = n_layers
+ self.gin_channels = gin_channels
+
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
+ self.enc = WN(
+ hidden_channels,
+ kernel_size,
+ dilation_rate,
+ n_layers,
+ gin_channels=gin_channels,
+ )
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
+
+ def __call__(
+ self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+ return super().__call__(x, x_lengths, g=g)
+
+ def forward(
+ self, x: torch.Tensor, x_lengths: torch.Tensor, g: Optional[torch.Tensor] = None
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+ x_mask = torch.unsqueeze(
+ sequence_mask(x_lengths, x.size(2)),
+ 1,
+ ).to(x.dtype)
+ x = self.pre(x) * x_mask
+ x = self.enc(x, x_mask, g=g)
+ stats = self.proj(x) * x_mask
+ m, logs = torch.split(stats, self.out_channels, dim=1)
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
+ return z, m, logs, x_mask
+
+ def remove_weight_norm(self):
+ self.enc.remove_weight_norm()
+
+ def __prepare_scriptable__(self):
+ for hook in self.enc._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.enc)
+ return self
diff --git a/rvc/layers/generators.py b/rvc/layers/generators.py
new file mode 100644
index 0000000000000000000000000000000000000000..945a41cbb694eec8361fcc77d1a4a1ad124e0118
--- /dev/null
+++ b/rvc/layers/generators.py
@@ -0,0 +1,228 @@
+from typing import Optional, List, Tuple
+
+import torch
+from torch import nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from .residuals import ResBlock1, ResBlock2, LRELU_SLOPE
+from .utils import call_weight_data_normal_if_Conv
+
+
+class Generator(torch.nn.Module):
+ def __init__(
+ self,
+ initial_channel: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ gin_channels: int = 0,
+ ):
+ super(Generator, self).__init__()
+ self.num_kernels = len(resblock_kernel_sizes)
+ self.num_upsamples = len(upsample_rates)
+
+ self.conv_pre = Conv1d(
+ initial_channel, upsample_initial_channel, 7, 1, padding=3
+ )
+
+ self.ups = nn.ModuleList()
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+ self.ups.append(
+ weight_norm(
+ ConvTranspose1d(
+ upsample_initial_channel // (2**i),
+ upsample_initial_channel // (2 ** (i + 1)),
+ k,
+ u,
+ padding=(k - u) // 2,
+ )
+ )
+ )
+
+ self.resblocks = nn.ModuleList()
+ resblock_module = ResBlock1 if resblock == "1" else ResBlock2
+ for i in range(len(self.ups)):
+ ch = upsample_initial_channel // (2 ** (i + 1))
+ for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes):
+ self.resblocks.append(resblock_module(ch, k, d))
+
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+ self.ups.apply(call_weight_data_normal_if_Conv)
+
+ if gin_channels != 0:
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ n_res: Optional[int] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, g=g, n_res=n_res)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ n_res: Optional[int] = None,
+ ):
+ if n_res is not None:
+ n = int(n_res)
+ if n != x.shape[-1]:
+ x = F.interpolate(x, size=n, mode="linear")
+
+ x = self.conv_pre(x)
+ if g is not None:
+ x = x + self.cond(g)
+
+ for i in range(self.num_upsamples):
+ x = F.leaky_relu(x, LRELU_SLOPE)
+ x = self.ups[i](x)
+ n = i * self.num_kernels
+ xs = self.resblocks[n](x)
+ for j in range(1, self.num_kernels):
+ xs += self.resblocks[n + j](x)
+ x = xs / self.num_kernels
+
+ x = F.leaky_relu(x)
+ x = self.conv_post(x)
+ x = torch.tanh(x)
+
+ return x
+
+ def __prepare_scriptable__(self):
+ for l in self.ups:
+ for hook in l._forward_pre_hooks.values():
+ # The hook we want to remove is an instance of WeightNorm class, so
+ # normally we would do `if isinstance(...)` but this class is not accessible
+ # because of shadowing, so we check the module name directly.
+ # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+
+ for l in self.resblocks:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ return self
+
+ def remove_weight_norm(self):
+ for l in self.ups:
+ remove_weight_norm(l)
+ for l in self.resblocks:
+ l.remove_weight_norm()
+
+
+class SineGenerator(torch.nn.Module):
+ """Definition of sine generator
+ SineGenerator(samp_rate, harmonic_num = 0,
+ sine_amp = 0.1, noise_std = 0.003,
+ voiced_threshold = 0,
+ flag_for_pulse=False)
+ samp_rate: sampling rate in Hz
+ harmonic_num: number of harmonic overtones (default 0)
+ sine_amp: amplitude of sine-wavefrom (default 0.1)
+ noise_std: std of Gaussian noise (default 0.003)
+ voiced_thoreshold: F0 threshold for U/V classification (default 0)
+ flag_for_pulse: this SinGen is used inside PulseGen (default False)
+ Note: when flag_for_pulse is True, the first time step of a voiced
+ segment is always sin(torch.pi) or cos(0)
+ """
+
+ def __init__(
+ self,
+ samp_rate: int,
+ harmonic_num: int = 0,
+ sine_amp: float = 0.1,
+ noise_std: float = 0.003,
+ voiced_threshold: int = 0,
+ ):
+ super(SineGenerator, self).__init__()
+ self.sine_amp = sine_amp
+ self.noise_std = noise_std
+ self.harmonic_num = harmonic_num
+ self.dim = harmonic_num + 1
+ self.sampling_rate = samp_rate
+ self.voiced_threshold = voiced_threshold
+
+ def __call__(
+ self, f0: torch.Tensor, upp: int
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+ return super().__call__(f0, upp)
+
+ def forward(
+ self, f0: torch.Tensor, upp: int
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+ """sine_tensor, uv = forward(f0)
+ input F0: tensor(batchsize=1, length, dim=1)
+ f0 for unvoiced steps should be 0
+ output sine_tensor: tensor(batchsize=1, length, dim)
+ output uv: tensor(batchsize=1, length, 1)
+ """
+ with torch.no_grad():
+ f0 = f0[:, None].transpose(1, 2)
+ f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
+ # fundamental component
+ f0_buf[:, :, 0] = f0[:, :, 0]
+ for idx in range(self.harmonic_num):
+ f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (
+ idx + 2
+ ) # idx + 2: the (idx+1)-th overtone, (idx+2)-th harmonic
+ rad_values = (
+ f0_buf / self.sampling_rate
+ ) % 1 ###%1意味着n_har的乘积无法后处理优化
+ rand_ini = torch.rand(
+ f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device
+ )
+ rand_ini[:, 0] = 0
+ rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
+ tmp_over_one = torch.cumsum(
+ rad_values, 1
+ ) # % 1 #####%1意味着后面的cumsum无法再优化
+ tmp_over_one *= upp
+ tmp_over_one: torch.Tensor = F.interpolate(
+ tmp_over_one.transpose(2, 1),
+ scale_factor=float(upp),
+ mode="linear",
+ align_corners=True,
+ ).transpose(2, 1)
+ rad_values: torch.Tensor = F.interpolate(
+ rad_values.transpose(2, 1), scale_factor=float(upp), mode="nearest"
+ ).transpose(
+ 2, 1
+ ) #######
+ tmp_over_one %= 1
+ tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
+ cumsum_shift = torch.zeros_like(rad_values)
+ cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
+ sine_waves = torch.sin(
+ torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * torch.pi
+ )
+ sine_waves = sine_waves * self.sine_amp
+ uv = self._f02uv(f0)
+ uv: torch.Tensor = F.interpolate(
+ uv.transpose(2, 1), scale_factor=float(upp), mode="nearest"
+ ).transpose(2, 1)
+ noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
+ noise = noise_amp * torch.randn_like(sine_waves)
+ sine_waves = sine_waves * uv + noise
+ return sine_waves, uv, noise
+
+ def _f02uv(self, f0):
+ # generate uv signal
+ uv = torch.ones_like(f0)
+ uv = uv * (f0 > self.voiced_threshold)
+ if uv.device.type == "privateuseone": # for DirectML
+ uv = uv.float()
+ return uv
diff --git a/rvc/layers/norms.py b/rvc/layers/norms.py
new file mode 100644
index 0000000000000000000000000000000000000000..9418035d2bc9b36545fc5a3652f3147408bc779f
--- /dev/null
+++ b/rvc/layers/norms.py
@@ -0,0 +1,148 @@
+from typing import Optional
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from .utils import activate_add_tanh_sigmoid_multiply
+
+
+class LayerNorm(nn.Module):
+ def __init__(self, channels: int, eps: float = 1e-5):
+ super(LayerNorm, self).__init__()
+ self.channels = channels
+ self.eps = eps
+
+ self.gamma = nn.Parameter(torch.ones(channels))
+ self.beta = nn.Parameter(torch.zeros(channels))
+
+ def forward(self, x: torch.Tensor):
+ x = x.transpose(1, -1)
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+ return x.transpose(1, -1)
+
+
+class WN(torch.nn.Module):
+ def __init__(
+ self,
+ hidden_channels: int,
+ kernel_size: int,
+ dilation_rate: int,
+ n_layers: int,
+ gin_channels: int = 0,
+ p_dropout: int = 0,
+ ):
+ super(WN, self).__init__()
+ assert kernel_size % 2 == 1
+ self.hidden_channels = hidden_channels
+ self.kernel_size = (kernel_size,)
+ self.dilation_rate = dilation_rate
+ self.n_layers = n_layers
+ self.gin_channels = gin_channels
+ self.p_dropout = float(p_dropout)
+
+ self.in_layers = torch.nn.ModuleList()
+ self.res_skip_layers = torch.nn.ModuleList()
+ self.drop = nn.Dropout(float(p_dropout))
+
+ if gin_channels != 0:
+ cond_layer = torch.nn.Conv1d(
+ gin_channels, 2 * hidden_channels * n_layers, 1
+ )
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
+
+ for i in range(n_layers):
+ dilation = dilation_rate**i
+ padding = int((kernel_size * dilation - dilation) / 2)
+ in_layer = torch.nn.Conv1d(
+ hidden_channels,
+ 2 * hidden_channels,
+ kernel_size,
+ dilation=dilation,
+ padding=padding,
+ )
+ in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
+ self.in_layers.append(in_layer)
+
+ # last one is not necessary
+ if i < n_layers - 1:
+ res_skip_channels = 2 * hidden_channels
+ else:
+ res_skip_channels = hidden_channels
+
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
+ self.res_skip_layers.append(res_skip_layer)
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, x_mask, g=g)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ output = torch.zeros_like(x)
+
+ if g is not None:
+ g = self.cond_layer(g)
+
+ for i, (in_layer, res_skip_layer) in enumerate(
+ zip(self.in_layers, self.res_skip_layers)
+ ):
+ x_in: torch.Tensor = in_layer(x)
+ if g is not None:
+ cond_offset = i * 2 * self.hidden_channels
+ g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
+ else:
+ g_l = torch.zeros_like(x_in)
+
+ acts = activate_add_tanh_sigmoid_multiply(x_in, g_l, self.hidden_channels)
+ acts: torch.Tensor = self.drop(acts)
+
+ res_skip_acts: torch.Tensor = res_skip_layer(acts)
+ if i < self.n_layers - 1:
+ res_acts = res_skip_acts[:, : self.hidden_channels, :]
+ x = (x + res_acts) * x_mask
+ output = output + res_skip_acts[:, self.hidden_channels :, :]
+ else:
+ output = output + res_skip_acts
+ return output * x_mask
+
+ def remove_weight_norm(self):
+ if self.gin_channels != 0:
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
+ for l in self.in_layers:
+ torch.nn.utils.remove_weight_norm(l)
+ for l in self.res_skip_layers:
+ torch.nn.utils.remove_weight_norm(l)
+
+ def __prepare_scriptable__(self):
+ if self.gin_channels != 0:
+ for hook in self.cond_layer._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
+ for l in self.in_layers:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ for l in self.res_skip_layers:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ return self
diff --git a/rvc/layers/nsf.py b/rvc/layers/nsf.py
new file mode 100644
index 0000000000000000000000000000000000000000..5e9e35ad77ae5ada6871741e8b13465610c73101
--- /dev/null
+++ b/rvc/layers/nsf.py
@@ -0,0 +1,216 @@
+from typing import Optional, List
+import math
+
+import torch
+from torch import nn
+from torch.nn import Conv1d, ConvTranspose1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from .generators import SineGenerator
+from .residuals import ResBlock1, ResBlock2, LRELU_SLOPE
+from .utils import call_weight_data_normal_if_Conv
+
+
+class SourceModuleHnNSF(torch.nn.Module):
+ """SourceModule for hn-nsf
+ SourceModule(sampling_rate, harmonic_num=0, sine_amp=0.1,
+ add_noise_std=0.003, voiced_threshod=0)
+ sampling_rate: sampling_rate in Hz
+ harmonic_num: number of harmonic above F0 (default: 0)
+ sine_amp: amplitude of sine source signal (default: 0.1)
+ add_noise_std: std of additive Gaussian noise (default: 0.003)
+ note that amplitude of noise in unvoiced is decided
+ by sine_amp
+ voiced_threshold: threhold to set U/V given F0 (default: 0)
+ Sine_source, noise_source = SourceModuleHnNSF(F0_sampled)
+ F0_sampled (batchsize, length, 1)
+ Sine_source (batchsize, length, 1)
+ noise_source (batchsize, length 1)
+ uv (batchsize, length, 1)
+ """
+
+ def __init__(
+ self,
+ sampling_rate: int,
+ harmonic_num: int = 0,
+ sine_amp: float = 0.1,
+ add_noise_std: float = 0.003,
+ voiced_threshod: int = 0,
+ ):
+ super(SourceModuleHnNSF, self).__init__()
+
+ self.sine_amp = sine_amp
+ self.noise_std = add_noise_std
+ # to produce sine waveforms
+ self.l_sin_gen = SineGenerator(
+ sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod
+ )
+ # to merge source harmonics into a single excitation
+ self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
+ self.l_tanh = torch.nn.Tanh()
+
+ def __call__(self, x: torch.Tensor, upp: int = 1) -> torch.Tensor:
+ return super().__call__(x, upp=upp)
+
+ def forward(self, x: torch.Tensor, upp: int = 1) -> torch.Tensor:
+ sine_wavs, _, _ = self.l_sin_gen(x, upp)
+ sine_wavs = sine_wavs.to(dtype=self.l_linear.weight.dtype)
+ sine_merge: torch.Tensor = self.l_tanh(self.l_linear(sine_wavs))
+ return sine_merge # , None, None # noise, uv
+
+
+class NSFGenerator(torch.nn.Module):
+ def __init__(
+ self,
+ initial_channel: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ gin_channels: int,
+ sr: int,
+ ):
+ super(NSFGenerator, self).__init__()
+ self.num_kernels = len(resblock_kernel_sizes)
+ self.num_upsamples = len(upsample_rates)
+
+ self.f0_upsamp = torch.nn.Upsample(scale_factor=math.prod(upsample_rates))
+ self.m_source = SourceModuleHnNSF(sampling_rate=sr, harmonic_num=0)
+ self.noise_convs = nn.ModuleList()
+ self.conv_pre = Conv1d(
+ initial_channel, upsample_initial_channel, 7, 1, padding=3
+ )
+ resblock = ResBlock1 if resblock == "1" else ResBlock2
+
+ self.ups = nn.ModuleList()
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+ c_cur = upsample_initial_channel // (2 ** (i + 1))
+ self.ups.append(
+ weight_norm(
+ ConvTranspose1d(
+ upsample_initial_channel // (2**i),
+ upsample_initial_channel // (2 ** (i + 1)),
+ k,
+ u,
+ padding=(k - u) // 2,
+ )
+ )
+ )
+ if i + 1 < len(upsample_rates):
+ stride_f0 = math.prod(upsample_rates[i + 1 :])
+ self.noise_convs.append(
+ Conv1d(
+ 1,
+ c_cur,
+ kernel_size=stride_f0 * 2,
+ stride=stride_f0,
+ padding=stride_f0 // 2,
+ )
+ )
+ else:
+ self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
+
+ self.resblocks = nn.ModuleList()
+ for i in range(len(self.ups)):
+ ch: int = upsample_initial_channel // (2 ** (i + 1))
+ for j, (k, d) in enumerate(
+ zip(resblock_kernel_sizes, resblock_dilation_sizes)
+ ):
+ self.resblocks.append(resblock(ch, k, d))
+
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
+ self.ups.apply(call_weight_data_normal_if_Conv)
+
+ if gin_channels != 0:
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
+
+ self.upp = math.prod(upsample_rates)
+
+ self.lrelu_slope = LRELU_SLOPE
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ f0: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ n_res: Optional[int] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, f0, g=g, n_res=n_res)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ f0: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ n_res: Optional[int] = None,
+ ) -> torch.Tensor:
+ har_source = self.m_source(f0, self.upp)
+ har_source = har_source.transpose(1, 2)
+
+ if n_res is not None:
+ n_res = int(n_res)
+ if n_res * self.upp != har_source.shape[-1]:
+ har_source = F.interpolate(
+ har_source, size=n_res * self.upp, mode="linear"
+ )
+ if n_res != x.shape[-1]:
+ x = F.interpolate(x, size=n_res, mode="linear")
+
+ x = self.conv_pre(x)
+ if g is not None:
+ x = x + self.cond(g)
+ # torch.jit.script() does not support direct indexing of torch modules
+ # That's why I wrote this
+ for i, (ups, noise_convs) in enumerate(zip(self.ups, self.noise_convs)):
+ if i < self.num_upsamples:
+ x = F.leaky_relu(x, self.lrelu_slope)
+ x = ups(x)
+ x_source = noise_convs(har_source)
+ x = x + x_source
+ xs: Optional[torch.Tensor] = None
+ l = [i * self.num_kernels + j for j in range(self.num_kernels)]
+ for j, resblock in enumerate(self.resblocks):
+ if j in l:
+ if xs is None:
+ xs = resblock(x)
+ else:
+ xs += resblock(x)
+ # This assertion cannot be ignored! \
+ # If ignored, it will cause torch.jit.script() compilation errors
+ assert isinstance(xs, torch.Tensor)
+ x = xs / self.num_kernels
+ x = F.leaky_relu(x)
+ x = self.conv_post(x)
+ x = torch.tanh(x)
+
+ return x
+
+ def remove_weight_norm(self):
+ for l in self.ups:
+ remove_weight_norm(l)
+ for l in self.resblocks:
+ l.remove_weight_norm()
+
+ def __prepare_scriptable__(self):
+ for l in self.ups:
+ for hook in l._forward_pre_hooks.values():
+ # The hook we want to remove is an instance of WeightNorm class, so
+ # normally we would do `if isinstance(...)` but this class is not accessible
+ # because of shadowing, so we check the module name directly.
+ # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ for l in self.resblocks:
+ for hook in self.resblocks._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ return self
diff --git a/rvc/layers/residuals.py b/rvc/layers/residuals.py
new file mode 100644
index 0000000000000000000000000000000000000000..4a135ceba0f280a18d6e9dea1a07fb832f599bfc
--- /dev/null
+++ b/rvc/layers/residuals.py
@@ -0,0 +1,353 @@
+from typing import Optional, List, Tuple
+
+import torch
+from torch import nn
+from torch.nn import Conv1d
+from torch.nn import functional as F
+from torch.nn.utils import remove_weight_norm, weight_norm
+
+from .norms import WN
+from .utils import (
+ get_padding,
+ call_weight_data_normal_if_Conv,
+)
+
+LRELU_SLOPE = 0.1
+
+
+class ResBlock1(torch.nn.Module):
+ def __init__(
+ self,
+ channels: int,
+ kernel_size: int = 3,
+ dilation: List[int] = (1, 3, 5),
+ ):
+ super(ResBlock1, self).__init__()
+
+ self.convs1 = nn.ModuleList()
+ for d in dilation:
+ self.convs1.append(
+ weight_norm(
+ Conv1d(
+ channels,
+ channels,
+ kernel_size,
+ 1,
+ dilation=d,
+ padding=get_padding(kernel_size, d),
+ )
+ ),
+ )
+ self.convs1.apply(call_weight_data_normal_if_Conv)
+
+ self.convs2 = nn.ModuleList()
+ for _ in dilation:
+ self.convs2.append(
+ weight_norm(
+ Conv1d(
+ channels,
+ channels,
+ kernel_size,
+ 1,
+ dilation=1,
+ padding=get_padding(kernel_size, 1),
+ )
+ ),
+ )
+ self.convs2.apply(call_weight_data_normal_if_Conv)
+ self.lrelu_slope = LRELU_SLOPE
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ x_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, x_mask=x_mask)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ for c1, c2 in zip(self.convs1, self.convs2):
+ xt = F.leaky_relu(x, self.lrelu_slope)
+ if x_mask is not None:
+ xt = xt * x_mask
+ xt = c1(xt)
+ xt = F.leaky_relu(xt, self.lrelu_slope)
+ if x_mask is not None:
+ xt = xt * x_mask
+ xt = c2(xt)
+ x = xt + x
+ if x_mask is not None:
+ x = x * x_mask
+ return x
+
+ def remove_weight_norm(self):
+ for l in self.convs1:
+ remove_weight_norm(l)
+ for l in self.convs2:
+ remove_weight_norm(l)
+
+ def __prepare_scriptable__(self):
+ for l in self.convs1:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ for l in self.convs2:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ return self
+
+
+class ResBlock2(torch.nn.Module):
+ """
+ Actually this module is not used currently
+ because all configs specified "resblock": "1"
+ """
+
+ def __init__(
+ self,
+ channels: int,
+ kernel_size=3,
+ dilation: List[int] = (1, 3),
+ ):
+ super(ResBlock2, self).__init__()
+ self.convs = nn.ModuleList()
+ for d in dilation:
+ self.convs.append(
+ weight_norm(
+ Conv1d(
+ channels,
+ channels,
+ kernel_size,
+ 1,
+ dilation=d,
+ padding=get_padding(kernel_size, d),
+ )
+ ),
+ )
+ self.convs.apply(call_weight_data_normal_if_Conv)
+ self.lrelu_slope = LRELU_SLOPE
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ x_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ return super().__call__(x, x_mask=x_mask)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ for c in self.convs:
+ xt = F.leaky_relu(x, self.lrelu_slope)
+ if x_mask is not None:
+ xt = xt * x_mask
+ xt = c(xt)
+ x = xt + x
+ if x_mask is not None:
+ x = x * x_mask
+ return x
+
+ def remove_weight_norm(self):
+ for l in self.convs:
+ remove_weight_norm(l)
+
+ def __prepare_scriptable__(self):
+ for l in self.convs:
+ for hook in l._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(l)
+ return self
+
+
+class ResidualCouplingLayer(nn.Module):
+ def __init__(
+ self,
+ channels: int,
+ hidden_channels: int,
+ kernel_size: int,
+ dilation_rate: int,
+ n_layers: int,
+ p_dropout: int = 0,
+ gin_channels: int = 0,
+ mean_only: bool = False,
+ ):
+ assert channels % 2 == 0, "channels should be divisible by 2"
+ super(ResidualCouplingLayer, self).__init__()
+ self.channels = channels
+ self.hidden_channels = hidden_channels
+ self.kernel_size = kernel_size
+ self.dilation_rate = dilation_rate
+ self.n_layers = n_layers
+ self.half_channels = channels // 2
+ self.mean_only = mean_only
+
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
+ self.enc = WN(
+ hidden_channels,
+ kernel_size,
+ dilation_rate,
+ n_layers,
+ p_dropout=float(p_dropout),
+ gin_channels=gin_channels,
+ )
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
+ self.post.weight.data.zero_()
+ self.post.bias.data.zero_()
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ reverse: bool = False,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ return super().__call__(x, x_mask, g=g, reverse=reverse)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ reverse: bool = False,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
+ h = self.pre(x0) * x_mask
+ h = self.enc(h, x_mask, g=g)
+ stats = self.post(h) * x_mask
+ if not self.mean_only:
+ m, logs = torch.split(stats, [self.half_channels] * 2, 1)
+ else:
+ m = stats
+ logs = torch.zeros_like(m)
+
+ if not reverse:
+ x1 = m + x1 * torch.exp(logs) * x_mask
+ x = torch.cat([x0, x1], 1)
+ logdet = torch.sum(logs, [1, 2])
+ return x, logdet
+
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
+ x = torch.cat([x0, x1], 1)
+ return x, torch.zeros([1])
+
+ def remove_weight_norm(self):
+ self.enc.remove_weight_norm()
+
+ def __prepare_scriptable__(self):
+ for hook in self.enc._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.enc)
+ return self
+
+
+class ResidualCouplingBlock(nn.Module):
+ class Flip(nn.Module):
+ """
+ torch.jit.script() Compiled functions
+ can't take variable number of arguments or
+ use keyword-only arguments with defaults
+ """
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ reverse: bool = False,
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]:
+ x = torch.flip(x, [1])
+ if not reverse:
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
+ return x, logdet
+ else:
+ return x, torch.zeros([1], device=x.device)
+
+ def __init__(
+ self,
+ channels: int,
+ hidden_channels: int,
+ kernel_size: int,
+ dilation_rate: int,
+ n_layers: int,
+ n_flows: int = 4,
+ gin_channels: int = 0,
+ ):
+ super(ResidualCouplingBlock, self).__init__()
+ self.channels = channels
+ self.hidden_channels = hidden_channels
+ self.kernel_size = kernel_size
+ self.dilation_rate = dilation_rate
+ self.n_layers = n_layers
+ self.n_flows = n_flows
+ self.gin_channels = gin_channels
+
+ self.flows = nn.ModuleList()
+ for _ in range(n_flows):
+ self.flows.append(
+ ResidualCouplingLayer(
+ channels,
+ hidden_channels,
+ kernel_size,
+ dilation_rate,
+ n_layers,
+ gin_channels=gin_channels,
+ mean_only=True,
+ )
+ )
+ self.flows.append(self.Flip())
+
+ def __call__(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ reverse: bool = False,
+ ) -> torch.Tensor:
+ return super().__call__(x, x_mask, g=g, reverse=reverse)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ x_mask: torch.Tensor,
+ g: Optional[torch.Tensor] = None,
+ reverse: bool = False,
+ ) -> torch.Tensor:
+ if not reverse:
+ for flow in self.flows:
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
+ else:
+ for flow in reversed(self.flows):
+ x, _ = flow.forward(x, x_mask, g=g, reverse=reverse)
+ return x
+
+ def remove_weight_norm(self):
+ for i in range(self.n_flows):
+ self.flows[i * 2].remove_weight_norm()
+
+ def __prepare_scriptable__(self):
+ for i in range(self.n_flows):
+ for hook in self.flows[i * 2]._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.flows[i * 2])
+ return self
diff --git a/rvc/layers/synthesizers.py b/rvc/layers/synthesizers.py
new file mode 100644
index 0000000000000000000000000000000000000000..474781ed86f65ea284a9f5068c39121ea24e6c89
--- /dev/null
+++ b/rvc/layers/synthesizers.py
@@ -0,0 +1,396 @@
+from typing import Optional, List, Union
+
+import torch
+from torch import nn
+
+
+from .encoders import TextEncoder, PosteriorEncoder
+from .generators import Generator
+from .nsf import NSFGenerator
+from .residuals import ResidualCouplingBlock
+from .utils import (
+ slice_on_last_dim,
+ rand_slice_segments_on_last_dim,
+)
+
+
+class SynthesizerTrnMsNSFsid(nn.Module):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr: Optional[Union[str, int]],
+ encoder_dim: int,
+ use_f0: bool,
+ ):
+ super().__init__()
+ if isinstance(sr, str):
+ sr = {
+ "32k": 32000,
+ "40k": 40000,
+ "48k": 48000,
+ }[sr]
+ self.spec_channels = spec_channels
+ self.inter_channels = inter_channels
+ self.hidden_channels = hidden_channels
+ self.filter_channels = filter_channels
+ self.n_heads = n_heads
+ self.n_layers = n_layers
+ self.kernel_size = kernel_size
+ self.p_dropout = float(p_dropout)
+ self.resblock = resblock
+ self.resblock_kernel_sizes = resblock_kernel_sizes
+ self.resblock_dilation_sizes = resblock_dilation_sizes
+ self.upsample_rates = upsample_rates
+ self.upsample_initial_channel = upsample_initial_channel
+ self.upsample_kernel_sizes = upsample_kernel_sizes
+ self.segment_size = segment_size
+ self.gin_channels = gin_channels
+ self.spk_embed_dim = spk_embed_dim
+
+ self.enc_p = TextEncoder(
+ encoder_dim,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ float(p_dropout),
+ f0=use_f0,
+ )
+ if use_f0:
+ self.dec = NSFGenerator(
+ inter_channels,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ gin_channels=gin_channels,
+ sr=sr,
+ )
+ else:
+ self.dec = Generator(
+ inter_channels,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ gin_channels=gin_channels,
+ )
+ self.enc_q = PosteriorEncoder(
+ spec_channels,
+ inter_channels,
+ hidden_channels,
+ 5,
+ 1,
+ 16,
+ gin_channels=gin_channels,
+ )
+ self.flow = ResidualCouplingBlock(
+ inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels
+ )
+ self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
+
+ def remove_weight_norm(self):
+ self.dec.remove_weight_norm()
+ self.flow.remove_weight_norm()
+ if hasattr(self, "enc_q"):
+ self.enc_q.remove_weight_norm()
+
+ def __prepare_scriptable__(self):
+ for hook in self.dec._forward_pre_hooks.values():
+ # The hook we want to remove is an instance of WeightNorm class, so
+ # normally we would do `if isinstance(...)` but this class is not accessible
+ # because of shadowing, so we check the module name directly.
+ # https://github.com/pytorch/pytorch/blob/be0ca00c5ce260eb5bcec3237357f7a30cc08983/torch/nn/utils/__init__.py#L3
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.dec)
+ for hook in self.flow._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.flow)
+ if hasattr(self, "enc_q"):
+ for hook in self.enc_q._forward_pre_hooks.values():
+ if (
+ hook.__module__ == "torch.nn.utils.weight_norm"
+ and hook.__class__.__name__ == "WeightNorm"
+ ):
+ torch.nn.utils.remove_weight_norm(self.enc_q)
+ return self
+
+ @torch.jit.ignore
+ def forward(
+ self,
+ phone: torch.Tensor,
+ phone_lengths: torch.Tensor,
+ y: torch.Tensor,
+ y_lengths: torch.Tensor,
+ ds: Optional[torch.Tensor] = None,
+ pitch: Optional[torch.Tensor] = None,
+ pitchf: Optional[torch.Tensor] = None,
+ ): # 这里ds是id，[bs,1]
+ # print(1,pitch.shape)#[bs,t]
+ g = self.emb_g(ds).unsqueeze(-1) # [b, 256, 1]##1是t，广播的
+ m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
+ z_p = self.flow(z, y_mask, g=g)
+ z_slice, ids_slice = rand_slice_segments_on_last_dim(
+ z, y_lengths, self.segment_size
+ )
+ if pitchf is not None:
+ pitchf = slice_on_last_dim(pitchf, ids_slice, self.segment_size)
+ o = self.dec(z_slice, pitchf, g=g)
+ else:
+ o = self.dec(z_slice, g=g)
+ return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
+
+ @torch.jit.export
+ def infer(
+ self,
+ phone: torch.Tensor,
+ phone_lengths: torch.Tensor,
+ sid: torch.Tensor,
+ pitch: Optional[torch.Tensor] = None,
+ pitchf: Optional[torch.Tensor] = None, # nsff0
+ skip_head: Optional[int] = None,
+ return_length: Optional[int] = None,
+ return_length2: Optional[int] = None,
+ ):
+ g = self.emb_g(sid).unsqueeze(-1)
+ if skip_head is not None and return_length is not None:
+ head = int(skip_head)
+ length = int(return_length)
+ flow_head = head - 24
+ if flow_head < 0:
+ flow_head = 0
+ dec_head = head - flow_head
+ m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths, flow_head)
+ z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+ z = self.flow(z_p, x_mask, g=g, reverse=True)
+ z = z[:, :, dec_head : dec_head + length]
+ x_mask = x_mask[:, :, dec_head : dec_head + length]
+ if pitchf is not None:
+ pitchf = pitchf[:, head : head + length]
+ else:
+ m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+ z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
+ z = self.flow(z_p, x_mask, g=g, reverse=True)
+ del z_p, m_p, logs_p
+ if pitchf is not None:
+ o = self.dec(
+ z * x_mask,
+ pitchf,
+ g=g,
+ n_res=return_length2,
+ )
+ else:
+ o = self.dec(z * x_mask, g=g, n_res=return_length2)
+ del x_mask, z
+ return o # , x_mask, (z, z_p, m_p, logs_p)
+
+
+class SynthesizerTrnMs256NSFsid(SynthesizerTrnMsNSFsid):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr: Union[str, int],
+ ):
+ super().__init__(
+ spec_channels,
+ segment_size,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ p_dropout,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ spk_embed_dim,
+ gin_channels,
+ sr,
+ 256,
+ True,
+ )
+
+
+class SynthesizerTrnMs768NSFsid(SynthesizerTrnMsNSFsid):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr: Union[str, int],
+ ):
+ super().__init__(
+ spec_channels,
+ segment_size,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ p_dropout,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ spk_embed_dim,
+ gin_channels,
+ sr,
+ 768,
+ True,
+ )
+
+
+class SynthesizerTrnMs256NSFsid_nono(SynthesizerTrnMsNSFsid):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr=None,
+ ):
+ super().__init__(
+ spec_channels,
+ segment_size,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ p_dropout,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ spk_embed_dim,
+ gin_channels,
+ 256,
+ False,
+ )
+
+
+class SynthesizerTrnMs768NSFsid_nono(SynthesizerTrnMsNSFsid):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr=None,
+ ):
+ super().__init__(
+ spec_channels,
+ segment_size,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ p_dropout,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ spk_embed_dim,
+ gin_channels,
+ 768,
+ False,
+ )
diff --git a/rvc/layers/transforms.py b/rvc/layers/transforms.py
new file mode 100644
index 0000000000000000000000000000000000000000..2cb2a366cce72fdcb3d6b83b9aadb02209e64133
--- /dev/null
+++ b/rvc/layers/transforms.py
@@ -0,0 +1,209 @@
+from typing import Optional
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+
+DEFAULT_MIN_BIN_WIDTH = 1e-3
+DEFAULT_MIN_BIN_HEIGHT = 1e-3
+DEFAULT_MIN_DERIVATIVE = 1e-3
+
+
+def piecewise_rational_quadratic_transform(
+ inputs: torch.Tensor,
+ unnormalized_widths: torch.Tensor,
+ unnormalized_heights: torch.Tensor,
+ unnormalized_derivatives: torch.Tensor,
+ inverse: bool = False,
+ tails: Optional[str] = None,
+ tail_bound: float = 1.0,
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+ min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+ if tails is None:
+ spline_fn = rational_quadratic_spline
+ spline_kwargs = {}
+ else:
+ spline_fn = unconstrained_rational_quadratic_spline
+ spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
+
+ outputs, logabsdet = spline_fn(
+ inputs=inputs,
+ unnormalized_widths=unnormalized_widths,
+ unnormalized_heights=unnormalized_heights,
+ unnormalized_derivatives=unnormalized_derivatives,
+ inverse=inverse,
+ min_bin_width=min_bin_width,
+ min_bin_height=min_bin_height,
+ min_derivative=min_derivative,
+ **spline_kwargs
+ )
+ return outputs, logabsdet
+
+
+def searchsorted(bin_locations, inputs, eps=1e-6):
+ bin_locations[..., -1] += eps
+ return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
+
+
+def unconstrained_rational_quadratic_spline(
+ inputs: torch.Tensor,
+ unnormalized_widths: torch.Tensor,
+ unnormalized_heights: torch.Tensor,
+ unnormalized_derivatives: torch.Tensor,
+ inverse: bool = False,
+ tails: str = "linear",
+ tail_bound: float = 1.0,
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+ min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+ inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
+ outside_interval_mask = ~inside_interval_mask
+
+ outputs = torch.zeros_like(inputs)
+ logabsdet = torch.zeros_like(inputs)
+
+ if tails == "linear":
+ unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
+ constant = np.log(np.exp(1 - min_derivative) - 1)
+ unnormalized_derivatives[..., 0] = constant
+ unnormalized_derivatives[..., -1] = constant
+
+ outputs[outside_interval_mask] = inputs[outside_interval_mask]
+ logabsdet[outside_interval_mask] = 0
+ else:
+ raise RuntimeError("{} tails are not implemented.".format(tails))
+
+ (
+ outputs[inside_interval_mask],
+ logabsdet[inside_interval_mask],
+ ) = rational_quadratic_spline(
+ inputs=inputs[inside_interval_mask],
+ unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
+ unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
+ unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
+ inverse=inverse,
+ left=-tail_bound,
+ right=tail_bound,
+ bottom=-tail_bound,
+ top=tail_bound,
+ min_bin_width=min_bin_width,
+ min_bin_height=min_bin_height,
+ min_derivative=min_derivative,
+ )
+
+ return outputs, logabsdet
+
+
+def rational_quadratic_spline(
+ inputs: torch.Tensor,
+ unnormalized_widths: torch.Tensor,
+ unnormalized_heights: torch.Tensor,
+ unnormalized_derivatives: torch.Tensor,
+ inverse: bool = False,
+ left=0.0,
+ right=1.0,
+ bottom=0.0,
+ top=1.0,
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
+ min_derivative=DEFAULT_MIN_DERIVATIVE,
+):
+ if torch.min(inputs) < left or torch.max(inputs) > right:
+ raise ValueError("Input to a transform is not within its domain")
+
+ num_bins = unnormalized_widths.shape[-1]
+
+ if min_bin_width * num_bins > 1.0:
+ raise ValueError("Minimal bin width too large for the number of bins")
+ if min_bin_height * num_bins > 1.0:
+ raise ValueError("Minimal bin height too large for the number of bins")
+
+ widths = F.softmax(unnormalized_widths, dim=-1)
+ widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
+ cumwidths = torch.cumsum(widths, dim=-1)
+ cumwidths = F.pad(cumwidths, pad=(1, 0), mode="constant", value=0.0)
+ cumwidths = (right - left) * cumwidths + left
+ cumwidths[..., 0] = left
+ cumwidths[..., -1] = right
+ widths = cumwidths[..., 1:] - cumwidths[..., :-1]
+
+ derivatives = min_derivative + F.softplus(unnormalized_derivatives)
+
+ heights = F.softmax(unnormalized_heights, dim=-1)
+ heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
+ cumheights = torch.cumsum(heights, dim=-1)
+ cumheights = F.pad(cumheights, pad=(1, 0), mode="constant", value=0.0)
+ cumheights = (top - bottom) * cumheights + bottom
+ cumheights[..., 0] = bottom
+ cumheights[..., -1] = top
+ heights = cumheights[..., 1:] - cumheights[..., :-1]
+
+ if inverse:
+ bin_idx = searchsorted(cumheights, inputs)[..., None]
+ else:
+ bin_idx = searchsorted(cumwidths, inputs)[..., None]
+
+ input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
+ input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
+
+ input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
+ delta = heights / widths
+ input_delta = delta.gather(-1, bin_idx)[..., 0]
+
+ input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
+ input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
+
+ input_heights = heights.gather(-1, bin_idx)[..., 0]
+
+ if inverse:
+ a = (inputs - input_cumheights) * (
+ input_derivatives + input_derivatives_plus_one - 2 * input_delta
+ ) + input_heights * (input_delta - input_derivatives)
+ b = input_heights * input_derivatives - (inputs - input_cumheights) * (
+ input_derivatives + input_derivatives_plus_one - 2 * input_delta
+ )
+ c = -input_delta * (inputs - input_cumheights)
+
+ discriminant = b.pow(2) - 4 * a * c
+ assert (discriminant >= 0).all()
+
+ root = (2 * c) / (-b - torch.sqrt(discriminant))
+ outputs = root * input_bin_widths + input_cumwidths
+
+ theta_one_minus_theta = root * (1 - root)
+ denominator = input_delta + (
+ (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+ * theta_one_minus_theta
+ )
+ derivative_numerator = input_delta.pow(2) * (
+ input_derivatives_plus_one * root.pow(2)
+ + 2 * input_delta * theta_one_minus_theta
+ + input_derivatives * (1 - root).pow(2)
+ )
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+ return outputs, -logabsdet
+ else:
+ theta = (inputs - input_cumwidths) / input_bin_widths
+ theta_one_minus_theta = theta * (1 - theta)
+
+ numerator = input_heights * (
+ input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta
+ )
+ denominator = input_delta + (
+ (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
+ * theta_one_minus_theta
+ )
+ outputs = input_cumheights + numerator / denominator
+
+ derivative_numerator = input_delta.pow(2) * (
+ input_derivatives_plus_one * theta.pow(2)
+ + 2 * input_delta * theta_one_minus_theta
+ + input_derivatives * (1 - theta).pow(2)
+ )
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
+
+ return outputs, logabsdet
diff --git a/rvc/layers/utils.py b/rvc/layers/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..418578bae1cc17317eb8f05b7595f05fe7fb2676
--- /dev/null
+++ b/rvc/layers/utils.py
@@ -0,0 +1,82 @@
+from typing import List, Optional, Tuple, Iterator
+
+import torch
+
+
+def call_weight_data_normal_if_Conv(m: torch.nn.Module):
+ classname = m.__class__.__name__
+ if classname.find("Conv") != -1:
+ mean = 0.0
+ std = 0.01
+ m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size: int, dilation=1) -> int:
+ return int((kernel_size * dilation - dilation) / 2)
+
+
+def slice_on_last_dim(
+ x: torch.Tensor,
+ start_indices: List[int],
+ segment_size=4,
+) -> torch.Tensor:
+ new_shape = [*x.shape]
+ new_shape[-1] = segment_size
+ ret = torch.empty(new_shape, device=x.device)
+ for i in range(x.size(0)):
+ idx_str = start_indices[i]
+ idx_end = idx_str + segment_size
+ ret[i, ..., :] = x[i, ..., idx_str:idx_end]
+ return ret
+
+
+def rand_slice_segments_on_last_dim(
+ x: torch.Tensor,
+ x_lengths: int = None,
+ segment_size=4,
+) -> Tuple[torch.Tensor, List[int]]:
+ b, _, t = x.size()
+ if x_lengths is None:
+ x_lengths = t
+ ids_str_max = x_lengths - segment_size + 1
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
+ ret = slice_on_last_dim(x, ids_str, segment_size)
+ return ret, ids_str
+
+
+@torch.jit.script
+def activate_add_tanh_sigmoid_multiply(
+ input_a: torch.Tensor, input_b: torch.Tensor, n_channels: int
+) -> torch.Tensor:
+ in_act = input_a + input_b
+ t_act = torch.tanh(in_act[:, :n_channels, :])
+ s_act = torch.sigmoid(in_act[:, n_channels:, :])
+ acts = t_act * s_act
+ return acts
+
+
+def sequence_mask(
+ length: torch.Tensor,
+ max_length: Optional[int] = None,
+) -> torch.BoolTensor:
+ if max_length is None:
+ max_length = int(length.max())
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
+ return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def total_grad_norm(
+ parameters: Iterator[torch.nn.Parameter],
+ norm_type: float = 2.0,
+) -> float:
+ norm_type = float(norm_type)
+ total_norm = 0.0
+
+ for p in parameters:
+ if p.grad is None:
+ continue
+ param_norm = p.grad.data.norm(norm_type)
+ total_norm += float(param_norm.item()) ** norm_type
+ total_norm = total_norm ** (1.0 / norm_type)
+
+ return total_norm
diff --git a/rvc/onnx/__init__.py b/rvc/onnx/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..a6a5519c4b22b38d3d4f509427c3a94907b84071
--- /dev/null
+++ b/rvc/onnx/__init__.py
@@ -0,0 +1,2 @@
+from .infer import RVC
+from .exporter import export_onnx
diff --git a/rvc/onnx/exporter.py b/rvc/onnx/exporter.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b54bc44a4a1783ab2e0ea78d78494f8a6df4675
--- /dev/null
+++ b/rvc/onnx/exporter.py
@@ -0,0 +1,52 @@
+import torch
+
+from .synthesizer import SynthesizerTrnMsNSFsid
+
+
+def export_onnx(from_cpkt_pth: str, to_onnx_pth: str) -> str:
+ cpt = torch.load(from_cpkt_pth, map_location="cpu")
+ cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+ vec_channels = 256 if cpt.get("version", "v1") == "v1" else 768
+
+ test_phone = torch.rand(1, 200, vec_channels) # hidden unit
+ test_phone_lengths = torch.tensor([200]).long() # hidden unit 长度（貌似没啥用）
+ test_pitch = torch.randint(size=(1, 200), low=5, high=255) # 基频（单位赫兹）
+ test_pitchf = torch.rand(1, 200) # nsf基频
+ test_ds = torch.LongTensor([0]) # 说话人ID
+ test_rnd = torch.rand(1, 192, 200) # 噪声（加入随机因子）
+
+ device = "cpu" # 导出时设备（不影响使用模型）
+
+ net_g = SynthesizerTrnMsNSFsid(
+ *cpt["config"], encoder_dim=vec_channels
+ ) # fp32导出（C++要支持fp16必须手动将内存重新排列所以暂时不用fp16）
+ net_g.load_state_dict(cpt["weight"], strict=False)
+ input_names = ["phone", "phone_lengths", "pitch", "pitchf", "ds", "rnd"]
+ output_names = [
+ "audio",
+ ]
+ # net_g.construct_spkmixmap() #多角色混合轨道导出
+ torch.onnx.export(
+ net_g,
+ (
+ test_phone.to(device),
+ test_phone_lengths.to(device),
+ test_pitch.to(device),
+ test_pitchf.to(device),
+ test_ds.to(device),
+ test_rnd.to(device),
+ ),
+ to_onnx_pth,
+ dynamic_axes={
+ "phone": [1],
+ "pitch": [1],
+ "pitchf": [1],
+ "rnd": [2],
+ },
+ do_constant_folding=False,
+ opset_version=17,
+ verbose=False,
+ input_names=input_names,
+ output_names=output_names,
+ )
+ return "Finished"
diff --git a/rvc/onnx/infer.py b/rvc/onnx/infer.py
new file mode 100644
index 0000000000000000000000000000000000000000..b6895cda783bb9d71323478a6713ef1d786d3c12
--- /dev/null
+++ b/rvc/onnx/infer.py
@@ -0,0 +1,146 @@
+import typing
+import os
+
+import librosa
+import numpy as np
+import onnxruntime
+
+from rvc.f0 import (
+ PM,
+ Harvest,
+ Dio,
+ F0Predictor,
+)
+
+
+class Model:
+ def __init__(
+ self,
+ path: typing.Union[str, bytes, os.PathLike],
+ device: typing.Literal["cpu", "cuda", "dml"] = "cpu",
+ ):
+ if device == "cpu":
+ providers = ["CPUExecutionProvider"]
+ elif device == "cuda":
+ providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+ elif device == "dml":
+ providers = ["DmlExecutionProvider"]
+ else:
+ raise RuntimeError("Unsportted Device")
+ self.model = onnxruntime.InferenceSession(path, providers=providers)
+
+
+class ContentVec(Model):
+ def __init__(
+ self,
+ vec_path: typing.Union[str, bytes, os.PathLike],
+ device: typing.Literal["cpu", "cuda", "dml"] = "cpu",
+ ):
+ super().__init__(vec_path, device)
+
+ def __call__(self, wav: np.ndarray[typing.Any, np.dtype]):
+ return self.forward(wav)
+
+ def forward(self, wav: np.ndarray[typing.Any, np.dtype]):
+ if wav.ndim == 2: # double channels
+ wav = wav.mean(-1)
+ assert wav.ndim == 1, wav.ndim
+ wav = np.expand_dims(np.expand_dims(wav, 0), 0)
+ onnx_input = {self.model.get_inputs()[0].name: wav}
+ logits = self.model.run(None, onnx_input)[0]
+ return logits.transpose(0, 2, 1)
+
+
+predictors: typing.Dict[str, F0Predictor] = {
+ "pm": PM,
+ "harvest": Harvest,
+ "dio": Dio,
+}
+
+
+def get_f0_predictor(
+ f0_method: str, hop_length: int, sampling_rate: int
+) -> F0Predictor:
+ return predictors[f0_method](hop_length=hop_length, sampling_rate=sampling_rate)
+
+
+class RVC(Model):
+ def __init__(
+ self,
+ model_path: typing.Union[str, bytes, os.PathLike],
+ hop_len=512,
+ vec_path: typing.Union[str, bytes, os.PathLike] = "vec-768-layer-12.onnx",
+ device: typing.Literal["cpu", "cuda", "dml"] = "cpu",
+ ):
+ super().__init__(model_path, device)
+ self.vec_model = ContentVec(vec_path, device)
+ self.hop_len = hop_len
+
+ def infer(
+ self,
+ wav: np.ndarray[typing.Any, np.dtype],
+ wav_sr: int,
+ model_sr: int = 40000,
+ sid: int = 0,
+ f0_method="dio",
+ f0_up_key=0,
+ ) -> np.ndarray[typing.Any, np.dtype[np.int16]]:
+ f0_min = 50
+ f0_max = 1100
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+ f0_predictor = get_f0_predictor(
+ f0_method,
+ self.hop_len,
+ model_sr,
+ )
+ org_length = len(wav)
+ if org_length / wav_sr > 50.0:
+ raise RuntimeError("wav max length exceeded")
+
+ hubert = self.vec_model(librosa.resample(wav, orig_sr=wav_sr, target_sr=16000))
+ hubert = np.repeat(hubert, 2, axis=2).transpose(0, 2, 1).astype(np.float32)
+ hubert_length = hubert.shape[1]
+
+ pitchf = f0_predictor.compute_f0(wav, hubert_length)
+ pitchf = pitchf * 2 ** (f0_up_key / 12)
+ pitch = pitchf.copy()
+ f0_mel = 1127 * np.log(1 + pitch / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+ f0_mel_max - f0_mel_min
+ ) + 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > 255] = 255
+ pitch = np.rint(f0_mel).astype(np.int64)
+
+ pitchf = pitchf.reshape(1, len(pitchf)).astype(np.float32)
+ pitch = pitch.reshape(1, len(pitch))
+ ds = np.array([sid]).astype(np.int64)
+
+ rnd = np.random.randn(1, 192, hubert_length).astype(np.float32)
+ hubert_length = np.array([hubert_length]).astype(np.int64)
+
+ out_wav = self.forward(hubert, hubert_length, pitch, pitchf, ds, rnd).squeeze()
+
+ out_wav = np.pad(out_wav, (0, 2 * self.hop_len), "constant")
+
+ return out_wav[0:org_length]
+
+ def forward(
+ self,
+ hubert: np.ndarray[typing.Any, np.dtype[np.float32]],
+ hubert_length: int,
+ pitch: np.ndarray[typing.Any, np.dtype[np.int64]],
+ pitchf: np.ndarray[typing.Any, np.dtype[np.float32]],
+ ds: np.ndarray[typing.Any, np.dtype[np.int64]],
+ rnd: np.ndarray[typing.Any, np.dtype[np.float32]],
+ ) -> np.ndarray[typing.Any, np.dtype[np.int16]]:
+ onnx_input = {
+ self.model.get_inputs()[0].name: hubert,
+ self.model.get_inputs()[1].name: hubert_length,
+ self.model.get_inputs()[2].name: pitch,
+ self.model.get_inputs()[3].name: pitchf,
+ self.model.get_inputs()[4].name: ds,
+ self.model.get_inputs()[5].name: rnd,
+ }
+ return (self.model.run(None, onnx_input)[0] * 32767).astype(np.int16)
diff --git a/rvc/onnx/synthesizer.py b/rvc/onnx/synthesizer.py
new file mode 100644
index 0000000000000000000000000000000000000000..e8bf5161c1e5f89623c5dc1a011ca7551f1b6fa3
--- /dev/null
+++ b/rvc/onnx/synthesizer.py
@@ -0,0 +1,80 @@
+from typing import List, Optional, Union
+
+import torch
+
+from rvc.layers.synthesizers import SynthesizerTrnMsNSFsid as SynthesizerBase
+
+
+class SynthesizerTrnMsNSFsid(SynthesizerBase):
+ def __init__(
+ self,
+ spec_channels: int,
+ segment_size: int,
+ inter_channels: int,
+ hidden_channels: int,
+ filter_channels: int,
+ n_heads: int,
+ n_layers: int,
+ kernel_size: int,
+ p_dropout: int,
+ resblock: str,
+ resblock_kernel_sizes: List[int],
+ resblock_dilation_sizes: List[List[int]],
+ upsample_rates: List[int],
+ upsample_initial_channel: int,
+ upsample_kernel_sizes: List[int],
+ spk_embed_dim: int,
+ gin_channels: int,
+ sr: Optional[Union[str, int]],
+ encoder_dim: int,
+ ):
+ super().__init__(
+ spec_channels,
+ segment_size,
+ inter_channels,
+ hidden_channels,
+ filter_channels,
+ n_heads,
+ n_layers,
+ kernel_size,
+ p_dropout,
+ resblock,
+ resblock_kernel_sizes,
+ resblock_dilation_sizes,
+ upsample_rates,
+ upsample_initial_channel,
+ upsample_kernel_sizes,
+ spk_embed_dim,
+ gin_channels,
+ sr,
+ encoder_dim,
+ True,
+ )
+ self.speaker_map = None
+
+ def remove_weight_norm(self):
+ self.dec.remove_weight_norm()
+ self.flow.remove_weight_norm()
+ self.enc_q.remove_weight_norm()
+
+ def construct_spkmixmap(self):
+ self.speaker_map = torch.zeros((self.n_speaker, 1, 1, self.gin_channels))
+ for i in range(self.n_speaker):
+ self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
+ self.speaker_map = self.speaker_map.unsqueeze(0)
+
+ def forward(self, phone, phone_lengths, pitch, nsff0, g, rnd, max_len=None):
+ if self.speaker_map is not None: # [N, S] * [S, B, 1, H]
+ g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1)) # [N, S, B, 1, 1]
+ g = g * self.speaker_map # [N, S, B, 1, H]
+ g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
+ g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
+ else:
+ g = g.unsqueeze(0)
+ g = self.emb_g(g).transpose(1, 2)
+
+ m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
+ z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
+ z = self.flow(z_p, x_mask, g=g, reverse=True)
+ o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
+ return o
diff --git a/rvc/synthesizer.py b/rvc/synthesizer.py
new file mode 100644
index 0000000000000000000000000000000000000000..bfc7c4d150d13f006c14fc3844ff667f731671a2
--- /dev/null
+++ b/rvc/synthesizer.py
@@ -0,0 +1,65 @@
+from collections import OrderedDict
+
+import torch
+
+from .layers.synthesizers import SynthesizerTrnMsNSFsid
+from .jit import load_inputs, export_jit_model, save_pickle
+
+
+def get_synthesizer(cpt: OrderedDict, device=torch.device("cpu")):
+ cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
+ if_f0 = cpt.get("f0", 1)
+ version = cpt.get("version", "v1")
+ if version == "v1":
+ encoder_dim = 256
+ elif version == "v2":
+ encoder_dim = 768
+ net_g = SynthesizerTrnMsNSFsid(
+ *cpt["config"],
+ encoder_dim=encoder_dim,
+ use_f0=if_f0 == 1,
+ )
+ del net_g.enc_q
+ net_g.load_state_dict(cpt["weight"], strict=False)
+ net_g = net_g.float()
+ net_g.eval().to(device)
+ net_g.remove_weight_norm()
+ return net_g, cpt
+
+
+def load_synthesizer(
+ pth_path: torch.serialization.FILE_LIKE, device=torch.device("cpu")
+):
+ return get_synthesizer(
+ torch.load(pth_path, map_location=torch.device("cpu")),
+ device,
+ )
+
+
+def synthesizer_jit_export(
+ model_path: str,
+ mode: str = "script",
+ inputs_path: str = None,
+ save_path: str = None,
+ device=torch.device("cpu"),
+ is_half=False,
+):
+ if not save_path:
+ save_path = model_path.rstrip(".pth")
+ save_path += ".half.jit" if is_half else ".jit"
+ if "cuda" in str(device) and ":" not in str(device):
+ device = torch.device("cuda:0")
+ from rvc.synthesizer import load_synthesizer
+
+ model, cpt = load_synthesizer(model_path, device)
+ assert isinstance(cpt, dict)
+ model.forward = model.infer
+ inputs = None
+ if mode == "trace":
+ inputs = load_inputs(inputs_path, device, is_half)
+ ckpt = export_jit_model(model, mode, inputs, device, is_half)
+ cpt.pop("weight")
+ cpt["model"] = ckpt["model"]
+ cpt["device"] = device
+ save_pickle(cpt, save_path)
+ return cpt
diff --git a/sha256.env b/sha256.env
new file mode 100644
index 0000000000000000000000000000000000000000..53911bc5a898b9fb26af6b68817bc99f8a9cb12b
--- /dev/null
+++ b/sha256.env
@@ -0,0 +1,39 @@
+sha256_hubert_base_pt = f54b40fd2802423a5643779c4861af1e9ee9c1564dc9d32f54f20b5ffba7db96
+sha256_rmvpe_pt = 6d62215f4306e3ca278246188607209f09af3dc77ed4232efdd069798c4ec193
+sha256_rmvpe_onnx = 5370e71ac80af8b4b7c793d27efd51fd8bf962de3a7ede0766dac0befa3660fd
+
+sha256_v1_D32k_pth = 2ab20645829460fdad0d3c44254f1ab53c32cae50c22a66c926ae5aa30abda6f
+sha256_v1_D40k_pth = 547f66dbbcd9023b9051ed244d12ab043ba8a4e854b154cc28761ac7c002909b
+sha256_v1_D48k_pth = 8cc013fa60ed9c3f902f5bd99f48c7e3b9352d763d4d3cd6bc241c37b0bfd9ad
+sha256_v1_G32k_pth = 81817645cde7ed2e2d83f23ef883f33dda564924b497e84d792743912eca4c23
+sha256_v1_G40k_pth = e428573bda1124b0ae0ae843fd8dcded6027d3993444790b3e9b0100938b2113
+sha256_v1_G48k_pth = 3862a67ea6313e8ffefc05cee6bee656ef3e089442e9ecf4a6618d60721f3e95
+sha256_v1_f0D32k_pth = 294db3087236e2c75260d6179056791c9231245daf5d0485545d9e54c4057c77
+sha256_v1_f0D40k_pth = 7d4f5a441594b470d67579958b2fd4c6b992852ded28ff9e72eda67abcebe423
+sha256_v1_f0D48k_pth = 1b84c8bf347ad1e539c842e8f2a4c36ecd9e7fb23c16041189e4877e9b07925c
+sha256_v1_f0G32k_pth = 285f524bf48bb692c76ad7bd0bc654c12bd9e5edeb784dddf7f61a789a608574
+sha256_v1_f0G40k_pth = 9115654aeef1995f7dd3c6fc4140bebbef0ca9760bed798105a2380a34299831
+sha256_v1_f0G48k_pth = 78bc9cab27e34bcfc194f93029374d871d8b3e663ddedea32a9709e894cc8fe8
+
+sha256_v2_D32k_pth = d8043378cc6619083d385f5a045de09b83fb3bf8de45c433ca863b71723ac3ca
+sha256_v2_D40k_pth = 471378e894e7191f89a94eda8288c5947b16bbe0b10c3f1f17efdb7a1d998242
+sha256_v2_D48k_pth = db01094a93c09868a278e03dafe8bb781bfcc1a5ba8df168c948bf9168c84d82
+sha256_v2_G32k_pth = 869b26a47f75168d6126f64ac39e6de5247017a8658cfd68aca600f7323efb9f
+sha256_v2_G40k_pth = a3843da7fde33db1dab176146c70d6c2df06eafe9457f4e3aa10024e9c6a4b69
+sha256_v2_G48k_pth = 2e2b1581a436d07a76b10b9d38765f64aa02836dc65c7dee1ce4140c11ea158b
+sha256_v2_f0D32k_pth = bd7134e7793674c85474d5145d2d982e3c5d8124fc7bb6c20f710ed65808fa8a
+sha256_v2_f0D40k_pth = 6b6ab091e70801b28e3f41f335f2fc5f3f35c75b39ae2628d419644ec2b0fa09
+sha256_v2_f0D48k_pth = 2269b73c7a4cf34da09aea99274dabf99b2ddb8a42cbfb065fb3c0aa9a2fc748
+sha256_v2_f0G32k_pth = 2332611297b8d88c7436de8f17ef5f07a2119353e962cd93cda5806d59a1133d
+sha256_v2_f0G40k_pth = 3b2c44035e782c4b14ddc0bede9e2f4a724d025cd073f736d4f43708453adfcb
+sha256_v2_f0G48k_pth = b5d51f589cc3632d4eae36a315b4179397695042edc01d15312e1bddc2b764a4
+
+sha256_uvr5_HP2-人声vocals+非人声instrumentals_pth = 39796caa5db18d7f9382d8ac997ac967bfd85f7761014bb807d2543cc844ef05
+sha256_uvr5_HP2_all_vocals_pth = 39796caa5db18d7f9382d8ac997ac967bfd85f7761014bb807d2543cc844ef05
+sha256_uvr5_HP3_all_vocals_pth = 45e6b65199e781b4a6542002699be9f19cd3d1cb7d1558bc2bfbcd84674dfe28
+sha256_uvr5_HP5-主旋律人声vocals+其他instrumentals_pth = 5908891829634926119720241e8573d97cbeb8277110a7512bdb0bd7563258ee
+sha256_uvr5_HP5_only_main_vocal_pth = 5908891829634926119720241e8573d97cbeb8277110a7512bdb0bd7563258ee
+sha256_uvr5_VR-DeEchoAggressive_pth = 8c8fd1582f9aabc363e47af62ddb88df6cae7e064cae75bbf041a067a5e0aee2
+sha256_uvr5_VR-DeEchoDeReverb_pth = 01376dd2a571bf3cb9cced680732726d2d732609d09216a610b0d110f133febe
+sha256_uvr5_VR-DeEchoNormal_pth = 56aba59db3bcdd14a14464e62f3129698ecdea62eee0f003b9360923eb3ac79e
+sha256_uvr5_vocals_onnx = 233bb5c6aaa365e568659a0a81211746fa881f8f47f82d9e864fce1f7692db80
diff --git a/tools/checksum/main.go b/tools/checksum/main.go
new file mode 100644
index 0000000000000000000000000000000000000000..83594f2c2fa3335c3b937a3e9c4067a60708ec31
--- /dev/null
+++ b/tools/checksum/main.go
@@ -0,0 +1,38 @@
+package main
+
+import (
+ "crypto/sha256"
+ "encoding/hex"
+ "fmt"
+ "io"
+ "os"
+)
+
+func main() {
+ var buf [32]byte
+ h := sha256.New()
+ lst := make([]any, 0, 64)
+ for _, fname := range files {
+ f, err := os.Open("assets/" + fname)
+ if err != nil {
+ panic(err)
+ }
+ _, err = io.Copy(h, f)
+ if err != nil {
+ panic(err)
+ }
+ s := hex.EncodeToString(h.Sum(buf[:0]))
+ fmt.Println("sha256 of", fname, "=", s)
+ lst = append(lst, s)
+ h.Reset()
+ f.Close()
+ }
+ f, err := os.Create("sha256.env")
+ if err != nil {
+ panic(err)
+ }
+ _, err = fmt.Fprintf(f, envtmpl, lst...)
+ if err != nil {
+ panic(err)
+ }
+}
diff --git a/tools/checksum/tmpl.go b/tools/checksum/tmpl.go
new file mode 100644
index 0000000000000000000000000000000000000000..ab5b7a86ae3bffeb90470b642ec70059bb369cff
--- /dev/null
+++ b/tools/checksum/tmpl.go
@@ -0,0 +1,84 @@
+package main
+
+var files = [...]string{
+ "hubert/hubert_base.pt",
+ "rmvpe/rmvpe.pt",
+ "rmvpe/rmvpe.onnx",
+
+ "pretrained/D32k.pth",
+ "pretrained/D40k.pth",
+ "pretrained/D48k.pth",
+ "pretrained/G32k.pth",
+ "pretrained/G40k.pth",
+ "pretrained/G48k.pth",
+ "pretrained/f0D32k.pth",
+ "pretrained/f0D40k.pth",
+ "pretrained/f0D48k.pth",
+ "pretrained/f0G32k.pth",
+ "pretrained/f0G40k.pth",
+ "pretrained/f0G48k.pth",
+
+ "pretrained_v2/D32k.pth",
+ "pretrained_v2/D40k.pth",
+ "pretrained_v2/D48k.pth",
+ "pretrained_v2/G32k.pth",
+ "pretrained_v2/G40k.pth",
+ "pretrained_v2/G48k.pth",
+ "pretrained_v2/f0D32k.pth",
+ "pretrained_v2/f0D40k.pth",
+ "pretrained_v2/f0D48k.pth",
+ "pretrained_v2/f0G32k.pth",
+ "pretrained_v2/f0G40k.pth",
+ "pretrained_v2/f0G48k.pth",
+
+ "uvr5_weights/HP2-人声vocals+非人声instrumentals.pth",
+ "uvr5_weights/HP2_all_vocals.pth",
+ "uvr5_weights/HP3_all_vocals.pth",
+ "uvr5_weights/HP5-主旋律人声vocals+其他instrumentals.pth",
+ "uvr5_weights/HP5_only_main_vocal.pth",
+ "uvr5_weights/VR-DeEchoAggressive.pth",
+ "uvr5_weights/VR-DeEchoDeReverb.pth",
+ "uvr5_weights/VR-DeEchoNormal.pth",
+ "uvr5_weights/onnx_dereverb_By_FoxJoy/vocals.onnx",
+}
+
+const envtmpl = `sha256_hubert_base_pt = %s
+sha256_rmvpe_pt = %s
+sha256_rmvpe_onnx = %s
+
+sha256_v1_D32k_pth = %s
+sha256_v1_D40k_pth = %s
+sha256_v1_D48k_pth = %s
+sha256_v1_G32k_pth = %s
+sha256_v1_G40k_pth = %s
+sha256_v1_G48k_pth = %s
+sha256_v1_f0D32k_pth = %s
+sha256_v1_f0D40k_pth = %s
+sha256_v1_f0D48k_pth = %s
+sha256_v1_f0G32k_pth = %s
+sha256_v1_f0G40k_pth = %s
+sha256_v1_f0G48k_pth = %s
+
+sha256_v2_D32k_pth = %s
+sha256_v2_D40k_pth = %s
+sha256_v2_D48k_pth = %s
+sha256_v2_G32k_pth = %s
+sha256_v2_G40k_pth = %s
+sha256_v2_G48k_pth = %s
+sha256_v2_f0D32k_pth = %s
+sha256_v2_f0D40k_pth = %s
+sha256_v2_f0D48k_pth = %s
+sha256_v2_f0G32k_pth = %s
+sha256_v2_f0G40k_pth = %s
+sha256_v2_f0G48k_pth = %s
+
+sha256_uvr5_HP2-人声vocals+非人声instrumentals_pth = %s
+sha256_uvr5_HP2_all_vocals_pth = %s
+sha256_uvr5_HP3_all_vocals_pth = %s
+sha256_uvr5_HP5-主旋律人声vocals+其他instrumentals_pth = %s
+sha256_uvr5_HP5_only_main_vocal_pth = %s
+sha256_uvr5_VR-DeEchoAggressive_pth = %s
+sha256_uvr5_VR-DeEchoDeReverb_pth = %s
+sha256_uvr5_VR-DeEchoNormal_pth = %s
+sha256_uvr5_vocals_onnx = %s
+`
diff --git a/tools/cmd/calc_rvc_model_similarity.py b/tools/cmd/calc_rvc_model_similarity.py
new file mode 100644
index 0000000000000000000000000000000000000000..42496e088e51dc5162d0714470c2226f696e260c
--- /dev/null
+++ b/tools/cmd/calc_rvc_model_similarity.py
@@ -0,0 +1,96 @@
+# This code references https://huggingface.co/JosephusCheung/ASimilarityCalculatior/blob/main/qwerty.py
+# Fill in the path of the model to be queried and the root directory of the reference models, and this script will return the similarity between the model to be queried and all reference models.
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def cal_cross_attn(to_q, to_k, to_v, rand_input):
+ hidden_dim, embed_dim = to_q.shape
+ attn_to_q = nn.Linear(hidden_dim, embed_dim, bias=False)
+ attn_to_k = nn.Linear(hidden_dim, embed_dim, bias=False)
+ attn_to_v = nn.Linear(hidden_dim, embed_dim, bias=False)
+ attn_to_q.load_state_dict({"weight": to_q})
+ attn_to_k.load_state_dict({"weight": to_k})
+ attn_to_v.load_state_dict({"weight": to_v})
+
+ return torch.einsum(
+ "ik, jk -> ik",
+ F.softmax(
+ torch.einsum("ij, kj -> ik", attn_to_q(rand_input), attn_to_k(rand_input)),
+ dim=-1,
+ ),
+ attn_to_v(rand_input),
+ )
+
+
+def model_hash(filename):
+ try:
+ with open(filename, "rb") as file:
+ import hashlib
+
+ m = hashlib.sha256()
+
+ file.seek(0x100000)
+ m.update(file.read(0x10000))
+ return m.hexdigest()[0:8]
+ except FileNotFoundError:
+ return "NOFILE"
+
+
+def eval(model, n, input):
+ qk = f"enc_p.encoder.attn_layers.{n}.conv_q.weight"
+ uk = f"enc_p.encoder.attn_layers.{n}.conv_k.weight"
+ vk = f"enc_p.encoder.attn_layers.{n}.conv_v.weight"
+ atoq, atok, atov = model[qk][:, :, 0], model[uk][:, :, 0], model[vk][:, :, 0]
+
+ attn = cal_cross_attn(atoq, atok, atov, input)
+ return attn
+
+
+def main(path, root):
+ torch.manual_seed(114514)
+ model_a = torch.load(path, map_location="cpu")["weight"]
+
+ logger.info("Query:\t\t%s\t%s" % (path, model_hash(path)))
+
+ map_attn_a = {}
+ map_rand_input = {}
+ for n in range(6):
+ hidden_dim, embed_dim, _ = model_a[
+ f"enc_p.encoder.attn_layers.{n}.conv_v.weight"
+ ].shape
+ rand_input = torch.randn([embed_dim, hidden_dim])
+
+ map_attn_a[n] = eval(model_a, n, rand_input)
+ map_rand_input[n] = rand_input
+
+ del model_a
+
+ for name in sorted(list(os.listdir(root))):
+ path = "%s/%s" % (root, name)
+ model_b = torch.load(path, map_location="cpu")["weight"]
+
+ sims = []
+ for n in range(6):
+ attn_a = map_attn_a[n]
+ attn_b = eval(model_b, n, map_rand_input[n])
+
+ sim = torch.mean(torch.cosine_similarity(attn_a, attn_b))
+ sims.append(sim)
+
+ logger.info(
+ "Reference:\t%s\t%s\t%s"
+ % (path, model_hash(path), f"{torch.mean(torch.stack(sims)) * 1e2:.2f}%")
+ )
+
+
+if __name__ == "__main__":
+ query_path = r"assets\weights\mi v3.pth"
+ reference_root = r"assets\weights"
+ main(query_path, reference_root)
diff --git a/tools/cmd/infer-pm-index256.py b/tools/cmd/infer-pm-index256.py
new file mode 100644
index 0000000000000000000000000000000000000000..a9905548d31360585c78f7f6fba88def4059b815
--- /dev/null
+++ b/tools/cmd/infer-pm-index256.py
@@ -0,0 +1,203 @@
+"""
+
+对源特征进行检索
+"""
+
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import parselmouth
+import torch
+
+os.environ["CUDA_VISIBLE_DEVICES"] = "0"
+# import torchcrepe
+from time import time as ttime
+
+# import pyworld
+import librosa
+import numpy as np
+import soundfile as sf
+import torch.nn.functional as F
+from fairseq import checkpoint_utils
+
+# from models import SynthesizerTrn256#hifigan_nonsf
+# from lib.infer_pack.models import SynthesizerTrn256NSF as SynthesizerTrn256#hifigan_nsf
+from rvc.layers.synthesizers import (
+ SynthesizerTrnMs256NSFsid as SynthesizerTrn256,
+) # hifigan_nsf
+from scipy.io import wavfile
+
+# from lib.infer_pack.models import SynthesizerTrnMs256NSFsid_sim as SynthesizerTrn256#hifigan_nsf
+# from models import SynthesizerTrn256NSFsim as SynthesizerTrn256#hifigan_nsf
+# from models import SynthesizerTrn256NSFsimFlow as SynthesizerTrn256#hifigan_nsf
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model_path = r"E:\codes\py39\vits_vc_gpu_train\assets\hubert\hubert_base.pt" #
+logger.info("Load model(s) from {}".format(model_path))
+models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+ [model_path],
+ suffix="",
+)
+model = models[0]
+model = model.to(device)
+model = model.half()
+model.eval()
+
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],183,256,is_half=True)#hifigan#512#256
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],109,256,is_half=True)#hifigan#512#256
+net_g = SynthesizerTrn256(
+ 1025,
+ 32,
+ 192,
+ 192,
+ 768,
+ 2,
+ 6,
+ 3,
+ 0,
+ "1",
+ [3, 7, 11],
+ [[1, 3, 5], [1, 3, 5], [1, 3, 5]],
+ [10, 10, 2, 2],
+ 512,
+ [16, 16, 4, 4],
+ 183,
+ 256,
+ is_half=True,
+) # hifigan#512#256#no_dropout
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,3,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2,2],512,[16,16,4,4],0)#ts3
+# net_g = SynthesizerTrn256(1025,32,192,192,768,2,6,3,0.1,"1", [3,7,11],[[1,3,5], [1,3,5], [1,3,5]],[10,10,2],512,[16,16,4],0)#hifigan-ps-sr
+#
+# net_g = SynthesizerTrn(1025, 32, 192, 192, 768, 2, 6, 3, 0.1, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [5,5], 512, [15,15], 0)#ms
+# net_g = SynthesizerTrn(1025, 32, 192, 192, 768, 2, 6, 3, 0.1, "1", [3, 7, 11], [[1, 3, 5], [1, 3, 5], [1, 3, 5]], [10,10], 512, [16,16], 0)#idwt2
+
+# weights=torch.load("infer/ft-mi_1k-noD.pt")
+# weights=torch.load("infer/ft-mi-freeze-vocoder-flow-enc_q_1k.pt")
+# weights=torch.load("infer/ft-mi-freeze-vocoder_true_1k.pt")
+# weights=torch.load("infer/ft-mi-sim1k.pt")
+weights = torch.load("infer/ft-mi-no_opt-no_dropout.pt")
+logger.debug(net_g.load_state_dict(weights, strict=True))
+
+net_g.eval().to(device)
+net_g.half()
+
+
+def get_f0(x, p_len, f0_up_key=0):
+ time_step = 160 / 16000 * 1000
+ f0_min = 50
+ f0_max = 1100
+ f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+ f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+
+ f0 = (
+ parselmouth.Sound(x, 16000)
+ .to_pitch_ac(
+ time_step=time_step / 1000,
+ voicing_threshold=0.6,
+ pitch_floor=f0_min,
+ pitch_ceiling=f0_max,
+ )
+ .selected_array["frequency"]
+ )
+
+ pad_size = (p_len - len(f0) + 1) // 2
+ if pad_size > 0 or p_len - len(f0) - pad_size > 0:
+ f0 = np.pad(f0, [[pad_size, p_len - len(f0) - pad_size]], mode="constant")
+ f0 *= pow(2, f0_up_key / 12)
+ f0bak = f0.copy()
+
+ f0_mel = 1127 * np.log(1 + f0 / 700)
+ f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (
+ f0_mel_max - f0_mel_min
+ ) + 1
+ f0_mel[f0_mel <= 1] = 1
+ f0_mel[f0_mel > 255] = 255
+ # f0_mel[f0_mel > 188] = 188
+ f0_coarse = np.rint(f0_mel).astype(np.int32)
+ return f0_coarse, f0bak
+
+
+import faiss
+
+index = faiss.read_index("infer/added_IVF512_Flat_mi_baseline_src_feat.index")
+big_npy = np.load("infer/big_src_feature_mi.npy")
+ta0 = ta1 = ta2 = 0
+for idx, name in enumerate(
+ [
+ "冬之花clip1.wav",
+ ]
+): ##
+ wav_path = "todo-songs/%s" % name #
+ f0_up_key = -2 #
+ audio, sampling_rate = sf.read(wav_path)
+ if len(audio.shape) > 1:
+ audio = librosa.to_mono(audio.transpose(1, 0))
+ if sampling_rate != 16000:
+ audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+
+ feats = torch.from_numpy(audio).float()
+ if feats.dim() == 2: # double channels
+ feats = feats.mean(-1)
+ assert feats.dim() == 1, feats.dim()
+ feats = feats.view(1, -1)
+ padding_mask = torch.BoolTensor(feats.shape).fill_(False)
+ inputs = {
+ "source": feats.half().to(device),
+ "padding_mask": padding_mask.to(device),
+ "output_layer": 9, # layer 9
+ }
+ if torch.cuda.is_available():
+ torch.cuda.synchronize()
+ t0 = ttime()
+ with torch.no_grad():
+ logits = model.extract_features(**inputs)
+ feats = model.final_proj(logits[0])
+
+ ####索引优化
+ npy = feats[0].cpu().numpy().astype("float32")
+ D, I = index.search(npy, 1)
+ feats = (
+ torch.from_numpy(big_npy[I.squeeze()].astype("float16")).unsqueeze(0).to(device)
+ )
+
+ feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
+ if torch.cuda.is_available():
+ torch.cuda.synchronize()
+ t1 = ttime()
+ # p_len = min(feats.shape[1],10000,pitch.shape[0])#太大了爆显存
+ p_len = min(feats.shape[1], 10000) #
+ pitch, pitchf = get_f0(audio, p_len, f0_up_key)
+ p_len = min(feats.shape[1], 10000, pitch.shape[0]) # 太大了爆显存
+ if torch.cuda.is_available():
+ torch.cuda.synchronize()
+ t2 = ttime()
+ feats = feats[:, :p_len, :]
+ pitch = pitch[:p_len]
+ pitchf = pitchf[:p_len]
+ p_len = torch.LongTensor([p_len]).to(device)
+ pitch = torch.LongTensor(pitch).unsqueeze(0).to(device)
+ sid = torch.LongTensor([0]).to(device)
+ pitchf = torch.FloatTensor(pitchf).unsqueeze(0).to(device)
+ with torch.no_grad():
+ audio = (
+ net_g.infer(feats, p_len, sid, pitch=pitch, pitchf=pitchf)[0, 0]
+ .data.cpu()
+ .float()
+ .numpy()
+ ) # nsf
+ if torch.cuda.is_available():
+ torch.cuda.synchronize()
+ t3 = ttime()
+ ta0 += t1 - t0
+ ta1 += t2 - t1
+ ta2 += t3 - t2
+ # wavfile.write("ft-mi_1k-index256-noD-%s.wav"%name, 40000, audio)##
+ # wavfile.write("ft-mi-freeze-vocoder-flow-enc_q_1k-%s.wav"%name, 40000, audio)##
+ # wavfile.write("ft-mi-sim1k-%s.wav"%name, 40000, audio)##
+ wavfile.write("ft-mi-no_opt-no_dropout-%s.wav" % name, 40000, audio) ##
+
+
+logger.debug("%.2fs %.2fs %.2fs", ta0, ta1, ta2) #
diff --git a/tools/cmd/infer_batch_rvc.py b/tools/cmd/infer_batch_rvc.py
new file mode 100644
index 0000000000000000000000000000000000000000..140497c3b77d7666ccd208c080097029cf712e74
--- /dev/null
+++ b/tools/cmd/infer_batch_rvc.py
@@ -0,0 +1,72 @@
+import argparse
+import os
+import sys
+
+print("Command-line arguments:", sys.argv)
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import sys
+
+import tqdm as tq
+from dotenv import load_dotenv
+from scipy.io import wavfile
+
+from configs import Config
+from infer.modules.vc import VC
+
+
+def arg_parse() -> tuple:
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--f0up_key", type=int, default=0)
+ parser.add_argument("--input_path", type=str, help="input path")
+ parser.add_argument("--index_path", type=str, help="index path")
+ parser.add_argument("--f0method", type=str, default="harvest", help="harvest or pm")
+ parser.add_argument("--opt_path", type=str, help="opt path")
+ parser.add_argument("--model_name", type=str, help="store in assets/weight_root")
+ parser.add_argument("--index_rate", type=float, default=0.66, help="index rate")
+ parser.add_argument("--device", type=str, help="device")
+ parser.add_argument("--is_half", type=bool, help="use half -> True")
+ parser.add_argument("--filter_radius", type=int, default=3, help="filter radius")
+ parser.add_argument("--resample_sr", type=int, default=0, help="resample sr")
+ parser.add_argument("--rms_mix_rate", type=float, default=1, help="rms mix rate")
+ parser.add_argument("--protect", type=float, default=0.33, help="protect")
+
+ args = parser.parse_args()
+ sys.argv = sys.argv[:1]
+
+ return args
+
+
+def main():
+ load_dotenv()
+ args = arg_parse()
+ config = Config()
+ config.device = args.device if args.device else config.device
+ config.is_half = args.is_half if args.is_half else config.is_half
+ vc = VC(config)
+ vc.get_vc(args.model_name)
+ audios = os.listdir(args.input_path)
+ for file in tq.tqdm(audios):
+ if file.endswith(".wav"):
+ file_path = os.path.join(args.input_path, file)
+ _, wav_opt = vc.vc_single(
+ 0,
+ file_path,
+ args.f0up_key,
+ None,
+ args.f0method,
+ args.index_path,
+ None,
+ args.index_rate,
+ args.filter_radius,
+ args.resample_sr,
+ args.rms_mix_rate,
+ args.protect,
+ )
+ out_path = os.path.join(args.opt_path, file)
+ wavfile.write(out_path, wav_opt[0], wav_opt[1])
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tools/cmd/infer_cli.py b/tools/cmd/infer_cli.py
new file mode 100644
index 0000000000000000000000000000000000000000..afd33a97364b762aa07be46dfd911e414da5d5cf
--- /dev/null
+++ b/tools/cmd/infer_cli.py
@@ -0,0 +1,67 @@
+import argparse
+import os
+import sys
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+from dotenv import load_dotenv
+from scipy.io import wavfile
+
+from configs import Config
+from infer.modules.vc import VC
+
+####
+# USAGE
+#
+# In your Terminal or CMD or whatever
+
+
+def arg_parse() -> tuple:
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--f0up_key", type=int, default=0)
+ parser.add_argument("--input_path", type=str, help="input path")
+ parser.add_argument("--index_path", type=str, help="index path")
+ parser.add_argument("--f0method", type=str, default="harvest", help="harvest or pm")
+ parser.add_argument("--opt_path", type=str, help="opt path")
+ parser.add_argument("--model_name", type=str, help="store in assets/weight_root")
+ parser.add_argument("--index_rate", type=float, default=0.66, help="index rate")
+ parser.add_argument("--device", type=str, help="device")
+ parser.add_argument("--is_half", type=bool, help="use half -> True")
+ parser.add_argument("--filter_radius", type=int, default=3, help="filter radius")
+ parser.add_argument("--resample_sr", type=int, default=0, help="resample sr")
+ parser.add_argument("--rms_mix_rate", type=float, default=1, help="rms mix rate")
+ parser.add_argument("--protect", type=float, default=0.33, help="protect")
+
+ args = parser.parse_args()
+ sys.argv = sys.argv[:1]
+
+ return args
+
+
+def main():
+ load_dotenv()
+ args = arg_parse()
+ config = Config()
+ config.device = args.device if args.device else config.device
+ config.is_half = args.is_half if args.is_half else config.is_half
+ vc = VC(config)
+ vc.get_vc(args.model_name)
+ _, wav_opt = vc.vc_single(
+ 0,
+ args.input_path,
+ args.f0up_key,
+ None,
+ args.f0method,
+ args.index_path,
+ None,
+ args.index_rate,
+ args.filter_radius,
+ args.resample_sr,
+ args.rms_mix_rate,
+ args.protect,
+ )
+ wavfile.write(args.opt_path, wav_opt[0], wav_opt[1])
+
+
+if __name__ == "__main__":
+ main()
diff --git a/tools/cmd/onnx/export.py b/tools/cmd/onnx/export.py
new file mode 100644
index 0000000000000000000000000000000000000000..77be81439c8869127df5b8816a594d2d306b86be
--- /dev/null
+++ b/tools/cmd/onnx/export.py
@@ -0,0 +1,3 @@
+from rvc.onnx import export_onnx
+
+export_onnx("pt/Justin Bieber.pth", "pt/TestRvc_Rvc.onnx")
diff --git a/tools/cmd/onnx/infer.py b/tools/cmd/onnx/infer.py
new file mode 100644
index 0000000000000000000000000000000000000000..413647303879667025ed4cd89dea8e05e45a5585
--- /dev/null
+++ b/tools/cmd/onnx/infer.py
@@ -0,0 +1,22 @@
+import soundfile
+import librosa
+
+from rvc.onnx import RVC
+
+hop_size = 512
+sampling_rate = 40000 # 采样率
+f0_up_key = 0 # 升降调
+sid = 0 # 角色ID
+f0_method = "dio" # F0提取算法
+model_path = "exported_model.onnx" # 模型的完整路径
+vec_path = "vec-256-layer-9.onnx" # 需要onnx的vec模型
+wav_path = "123.wav" # 输入路径或ByteIO实例
+out_path = "out.wav" # 输出路径或ByteIO实例
+
+model = RVC(model_path, vec_path=vec_path, hop_len=hop_size, device="cuda")
+
+wav, sr = librosa.load(wav_path, sr=sampling_rate)
+
+audio = model.infer(wav, sr, sampling_rate, sid, f0_method, f0_up_key)
+
+soundfile.write(out_path, audio, sampling_rate)
diff --git a/tools/cmd/train-index-v2.py b/tools/cmd/train-index-v2.py
new file mode 100644
index 0000000000000000000000000000000000000000..47c706c704440c363df3ea0689d91c8536385009
--- /dev/null
+++ b/tools/cmd/train-index-v2.py
@@ -0,0 +1,80 @@
+"""
+格式：直接cid为自带的index位；aid放不下了，通过字典来查，反正就5w个
+"""
+
+import os
+import traceback
+import logging
+
+logger = logging.getLogger(__name__)
+
+from multiprocessing import cpu_count
+
+import faiss
+import numpy as np
+from sklearn.cluster import MiniBatchKMeans
+
+# ###########如果是原始特征要先写save
+n_cpu = 0
+if n_cpu == 0:
+ n_cpu = cpu_count()
+inp_root = r"./logs/anz/3_feature768"
+npys = []
+listdir_res = list(os.listdir(inp_root))
+for name in sorted(listdir_res):
+ phone = np.load("%s/%s" % (inp_root, name))
+ npys.append(phone)
+big_npy = np.concatenate(npys, 0)
+big_npy_idx = np.arange(big_npy.shape[0])
+np.random.shuffle(big_npy_idx)
+big_npy = big_npy[big_npy_idx]
+logger.debug(big_npy.shape) # (6196072, 192)#fp32#4.43G
+if big_npy.shape[0] > 2e5:
+ # if(1):
+ info = "Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0]
+ logger.info(info)
+ try:
+ big_npy = (
+ MiniBatchKMeans(
+ n_clusters=10000,
+ verbose=True,
+ batch_size=256 * n_cpu,
+ compute_labels=False,
+ init="random",
+ )
+ .fit(big_npy)
+ .cluster_centers_
+ )
+ except:
+ info = traceback.format_exc()
+ logger.warning(info)
+
+np.save("tools/infer/big_src_feature_mi.npy", big_npy)
+
+##################train+add
+# big_npy=np.load("/bili-coeus/jupyter/jupyterhub-liujing04/vits_ch/inference_f0/big_src_feature_mi.npy")
+n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
+index = faiss.index_factory(768, "IVF%s,Flat" % n_ivf) # mi
+logger.info("Training...")
+index_ivf = faiss.extract_index_ivf(index) #
+index_ivf.nprobe = 1
+index.train(big_npy)
+faiss.write_index(
+ index, "tools/infer/trained_IVF%s_Flat_baseline_src_feat_v2.index" % (n_ivf)
+)
+logger.info("Adding...")
+batch_size_add = 8192
+for i in range(0, big_npy.shape[0], batch_size_add):
+ index.add(big_npy[i : i + batch_size_add])
+faiss.write_index(
+ index, "tools/infer/added_IVF%s_Flat_mi_baseline_src_feat.index" % (n_ivf)
+)
+"""
+大小（都是FP32）
+big_src_feature 2.95G
+ (3098036, 256)
+big_emb 4.43G
+ (6196072, 192)
+big_emb双倍是因为求特征要repeat后再加pitch
+
+"""
diff --git a/tools/cmd/train-index.py b/tools/cmd/train-index.py
new file mode 100644
index 0000000000000000000000000000000000000000..400adc0c949f09a1ac775073f9845ddac2bddf84
--- /dev/null
+++ b/tools/cmd/train-index.py
@@ -0,0 +1,43 @@
+"""
+格式：直接cid为自带的index位；aid放不下了，通过字典来查，反正就5w个
+"""
+
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import faiss
+import numpy as np
+
+# ###########如果是原始特征要先写save
+inp_root = r"E:\codes\py39\dataset\mi\2-co256"
+npys = []
+for name in sorted(list(os.listdir(inp_root))):
+ phone = np.load("%s/%s" % (inp_root, name))
+ npys.append(phone)
+big_npy = np.concatenate(npys, 0)
+logger.debug(big_npy.shape) # (6196072, 192)#fp32#4.43G
+np.save("infer/big_src_feature_mi.npy", big_npy)
+
+##################train+add
+# big_npy=np.load("/bili-coeus/jupyter/jupyterhub-liujing04/vits_ch/inference_f0/big_src_feature_mi.npy")
+logger.debug(big_npy.shape)
+index = faiss.index_factory(256, "IVF512,Flat") # mi
+logger.info("Training...")
+index_ivf = faiss.extract_index_ivf(index) #
+index_ivf.nprobe = 9
+index.train(big_npy)
+faiss.write_index(index, "infer/trained_IVF512_Flat_mi_baseline_src_feat.index")
+logger.info("Adding...")
+index.add(big_npy)
+faiss.write_index(index, "infer/added_IVF512_Flat_mi_baseline_src_feat.index")
+"""
+大小（都是FP32）
+big_src_feature 2.95G
+ (3098036, 256)
+big_emb 4.43G
+ (6196072, 192)
+big_emb双倍是因为求特征要repeat后再加pitch
+
+"""
diff --git a/tools/cmd/trans_weights.py b/tools/cmd/trans_weights.py
new file mode 100644
index 0000000000000000000000000000000000000000..a8ff3b0249b6f9c2f2748652e7563b6f28546ae0
--- /dev/null
+++ b/tools/cmd/trans_weights.py
@@ -0,0 +1,18 @@
+import pdb
+
+import torch
+
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-suc\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder-flow-enc_q\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-freeze-vocoder\G_1000.pth")["model"]#sim_nsf#
+# a=torch.load(r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-test\G_1000.pth")["model"]#sim_nsf#
+a = torch.load(
+ r"E:\codes\py39\vits_vc_gpu_train\logs\ft-mi-no_opt-no_dropout\G_1000.pth"
+)[
+ "model"
+] # sim_nsf#
+for key in a.keys():
+ a[key] = a[key].half()
+# torch.save(a,"ft-mi-freeze-vocoder_true_1k.pt")#
+# torch.save(a,"ft-mi-sim1k.pt")#
+torch.save(a, "ft-mi-no_opt-no_dropout.pt") #
diff --git a/tools/web/infer-only.py b/tools/web/infer-only.py
new file mode 100644
index 0000000000000000000000000000000000000000..a9d5d789b95070733b8e57765c080a9117405baa
--- /dev/null
+++ b/tools/web/infer-only.py
@@ -0,0 +1,174 @@
+import logging
+import os
+
+# os.system("wget -P cvec/ https://huggingface.co/lj1995/VoiceConversionWebUI/resolve/main/hubert_base.pt")
+import gradio as gr
+from dotenv import load_dotenv
+
+from configs import Config
+from i18n.i18n import I18nAuto
+from infer.modules.vc import VC
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+logging.getLogger("markdown_it").setLevel(logging.WARNING)
+logging.getLogger("urllib3").setLevel(logging.WARNING)
+logging.getLogger("matplotlib").setLevel(logging.WARNING)
+logger = logging.getLogger(__name__)
+
+i18n = I18nAuto()
+logger.info(i18n)
+
+load_dotenv()
+config = Config()
+vc = VC(config)
+
+weight_root = os.getenv("weight_root")
+weight_uvr5_root = os.getenv("weight_uvr5_root")
+index_root = os.getenv("index_root")
+names = []
+hubert_model = None
+for name in os.listdir(weight_root):
+ if name.endswith(".pth"):
+ names.append(name)
+index_paths = []
+for root, dirs, files in os.walk(index_root, topdown=False):
+ for name in files:
+ if name.endswith(".index") and "trained" not in name:
+ index_paths.append("%s/%s" % (root, name))
+
+
+app = gr.Blocks()
+with app:
+ with gr.Tabs():
+ with gr.TabItem("在线demo"):
+ gr.Markdown(
+ value="""
+ RVC 在线demo
+ """
+ )
+ sid = gr.Dropdown(label=i18n("Inferencing voice"), choices=sorted(names))
+ with gr.Column():
+ spk_item = gr.Slider(
+ minimum=0,
+ maximum=2333,
+ step=1,
+ label=i18n("Select Speaker/Singer ID"),
+ value=0,
+ visible=False,
+ interactive=True,
+ )
+ sid.change(fn=vc.get_vc, inputs=[sid], outputs=[spk_item])
+ gr.Markdown(
+ value=i18n(
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)"
+ )
+ )
+ vc_input3 = gr.Audio(label="上传音频（长度小于90秒）")
+ vc_transform0 = gr.Number(
+ label=i18n(
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)"
+ ),
+ value=0,
+ )
+ f0method0 = gr.Radio(
+ label=i18n(
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement"
+ ),
+ choices=["pm", "harvest", "crepe", "rmvpe"],
+ value="pm",
+ interactive=True,
+ )
+ filter_radius0 = gr.Slider(
+ minimum=0,
+ maximum=7,
+ label=i18n(
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness."
+ ),
+ value=3,
+ step=1,
+ interactive=True,
+ )
+ with gr.Column():
+ file_index1 = gr.Textbox(
+ label=i18n(
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown"
+ ),
+ value="",
+ interactive=False,
+ visible=False,
+ )
+ file_index2 = gr.Dropdown(
+ label=i18n("Auto-detect index path and select from the dropdown"),
+ choices=sorted(index_paths),
+ interactive=True,
+ )
+ index_rate1 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n("Feature searching ratio"),
+ value=0.88,
+ interactive=True,
+ )
+ resample_sr0 = gr.Slider(
+ minimum=0,
+ maximum=48000,
+ label=i18n(
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling"
+ ),
+ value=0,
+ step=1,
+ interactive=True,
+ )
+ rms_mix_rate0 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n(
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume"
+ ),
+ value=1,
+ interactive=True,
+ )
+ protect0 = gr.Slider(
+ minimum=0,
+ maximum=0.5,
+ label=i18n(
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy"
+ ),
+ value=0.33,
+ step=0.01,
+ interactive=True,
+ )
+ f0_file = gr.File(
+ label=i18n(
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation"
+ )
+ )
+ but0 = gr.Button(i18n("Convert"), variant="primary")
+ vc_output1 = gr.Textbox(label=i18n("Output information"))
+ vc_output2 = gr.Audio(
+ label=i18n(
+ "Export audio (click on the three dots in the lower right corner to download)"
+ )
+ )
+ but0.click(
+ vc.vc_single,
+ [
+ spk_item,
+ vc_input3,
+ vc_transform0,
+ f0_file,
+ f0method0,
+ file_index1,
+ file_index2,
+ # file_big_npy1,
+ index_rate1,
+ filter_radius0,
+ resample_sr0,
+ rms_mix_rate0,
+ protect0,
+ ],
+ [vc_output1, vc_output2],
+ )
+
+
+app.launch()
diff --git a/web.py b/web.py
new file mode 100644
index 0000000000000000000000000000000000000000..1c428107c426059296523c5aa7f68cc11f6a2e69
--- /dev/null
+++ b/web.py
@@ -0,0 +1,1713 @@
+import os
+import sys
+from dotenv import load_dotenv
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+load_dotenv()
+load_dotenv("sha256.env")
+
+if sys.platform == "darwin":
+ os.environ["PYTORCH_ENABLE_MPS_FALLBACK"] = "1"
+
+from infer.modules.vc import VC, show_info, hash_similarity
+from infer.modules.uvr5.modules import uvr
+from infer.lib.train.process_ckpt import (
+ change_info,
+ extract_small_model,
+ merge,
+)
+from i18n.i18n import I18nAuto
+from configs import Config
+from sklearn.cluster import MiniBatchKMeans
+import torch, platform
+import numpy as np
+import gradio as gr
+import faiss
+import pathlib
+import json
+from time import sleep
+from subprocess import Popen
+from random import shuffle
+import warnings
+import traceback
+import threading
+import shutil
+import logging
+
+
+logging.getLogger("numba").setLevel(logging.WARNING)
+logging.getLogger("httpx").setLevel(logging.WARNING)
+
+logger = logging.getLogger(__name__)
+
+tmp = os.path.join(now_dir, "TEMP")
+shutil.rmtree(tmp, ignore_errors=True)
+os.makedirs(tmp, exist_ok=True)
+os.makedirs(os.path.join(now_dir, "logs"), exist_ok=True)
+os.makedirs(os.path.join(now_dir, "assets/weights"), exist_ok=True)
+os.environ["TEMP"] = tmp
+warnings.filterwarnings("ignore")
+torch.manual_seed(114514)
+
+
+config = Config()
+vc = VC(config)
+
+if not config.nocheck:
+ from infer.lib.rvcmd import check_all_assets, download_all_assets
+
+ if not check_all_assets(update=config.update):
+ if config.update:
+ download_all_assets(tmpdir=tmp)
+ if not check_all_assets(update=config.update):
+ logging.error("counld not satisfy all assets needed.")
+ exit(1)
+
+if config.dml == True:
+
+ def forward_dml(ctx, x, scale):
+ ctx.scale = scale
+ res = x.clone().detach()
+ return res
+
+ import fairseq
+
+ fairseq.modules.grad_multiply.GradMultiply.forward = forward_dml
+
+i18n = I18nAuto()
+logger.info(i18n)
+# 判断是否有能用来训练和加速推理的N卡
+ngpu = torch.cuda.device_count()
+gpu_infos = []
+mem = []
+if_gpu_ok = False
+
+if torch.cuda.is_available() or ngpu != 0:
+ for i in range(ngpu):
+ gpu_name = torch.cuda.get_device_name(i)
+ if any(
+ value in gpu_name.upper()
+ for value in [
+ "10",
+ "16",
+ "20",
+ "30",
+ "40",
+ "A2",
+ "A3",
+ "A4",
+ "P4",
+ "A50",
+ "500",
+ "A60",
+ "70",
+ "80",
+ "90",
+ "M4",
+ "T4",
+ "TITAN",
+ "4060",
+ "L",
+ "6000",
+ ]
+ ):
+ # A10#A100#V100#A40#P40#M40#K80#A4500
+ if_gpu_ok = True # 至少有一张能用的N卡
+ gpu_infos.append("%s\t%s" % (i, gpu_name))
+ mem.append(
+ int(
+ torch.cuda.get_device_properties(i).total_memory
+ / 1024
+ / 1024
+ / 1024
+ + 0.4
+ )
+ )
+if if_gpu_ok and len(gpu_infos) > 0:
+ gpu_info = "\n".join(gpu_infos)
+ default_batch_size = min(mem) // 2
+else:
+ gpu_info = i18n(
+ "Unfortunately, there is no compatible GPU available to support your training."
+ )
+ default_batch_size = 1
+gpus = "-".join([i[0] for i in gpu_infos])
+
+
+weight_root = os.getenv("weight_root")
+weight_uvr5_root = os.getenv("weight_uvr5_root")
+index_root = os.getenv("index_root")
+outside_index_root = os.getenv("outside_index_root")
+
+names = []
+for name in os.listdir(weight_root):
+ if name.endswith(".pth"):
+ names.append(name)
+index_paths = []
+
+
+def lookup_indices(index_root):
+ global index_paths
+ for root, dirs, files in os.walk(index_root, topdown=False):
+ for name in files:
+ if name.endswith(".index") and "trained" not in name:
+ index_paths.append("%s/%s" % (root, name))
+
+
+lookup_indices(index_root)
+lookup_indices(outside_index_root)
+uvr5_names = []
+for name in os.listdir(weight_uvr5_root):
+ if name.endswith(".pth") or "onnx" in name:
+ uvr5_names.append(name.replace(".pth", ""))
+
+
+def change_choices():
+ names = []
+ for name in os.listdir(weight_root):
+ if name.endswith(".pth"):
+ names.append(name)
+ index_paths = []
+ for root, dirs, files in os.walk(index_root, topdown=False):
+ for name in files:
+ if name.endswith(".index") and "trained" not in name:
+ index_paths.append("%s/%s" % (root, name))
+ return {"choices": sorted(names), "__type__": "update"}, {
+ "choices": sorted(index_paths),
+ "__type__": "update",
+ }
+
+
+def clean():
+ return {"value": "", "__type__": "update"}
+
+
+def export_onnx(ModelPath, ExportedPath):
+ from rvc.onnx import export_onnx as eo
+
+ eo(ModelPath, ExportedPath)
+
+
+sr_dict = {
+ "32k": 32000,
+ "40k": 40000,
+ "48k": 48000,
+}
+
+
+def if_done(done, p):
+ while 1:
+ if p.poll() is None:
+ sleep(0.5)
+ else:
+ break
+ done[0] = True
+
+
+def if_done_multi(done, ps):
+ while 1:
+ # poll==None代表进程未结束
+ # 只要有一个进程未结束都不停
+ flag = 1
+ for p in ps:
+ if p.poll() is None:
+ flag = 0
+ sleep(0.5)
+ break
+ if flag == 1:
+ break
+ done[0] = True
+
+
+def preprocess_dataset(trainset_dir, exp_dir, sr, n_p):
+ sr = sr_dict[sr]
+ os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
+ f = open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "w")
+ f.close()
+ cmd = '"%s" infer/modules/train/preprocess.py "%s" %s %s "%s/logs/%s" %s %.1f' % (
+ config.python_cmd,
+ trainset_dir,
+ sr,
+ n_p,
+ now_dir,
+ exp_dir,
+ config.noparallel,
+ config.preprocess_per,
+ )
+ logger.info("Execute: " + cmd)
+ # , stdin=PIPE, stdout=PIPE,stderr=PIPE,cwd=now_dir
+ p = Popen(cmd, shell=True)
+ # 煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
+ done = [False]
+ threading.Thread(
+ target=if_done,
+ args=(
+ done,
+ p,
+ ),
+ ).start()
+ while 1:
+ with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
+ yield (f.read())
+ sleep(1)
+ if done[0]:
+ break
+ with open("%s/logs/%s/preprocess.log" % (now_dir, exp_dir), "r") as f:
+ log = f.read()
+ logger.info(log)
+ yield log
+
+
+# but2.click(extract_f0,[gpus6,np7,f0method8,if_f0_3,trainset_dir4],[info2])
+def extract_f0_feature(gpus, n_p, f0method, if_f0, exp_dir, version19, gpus_rmvpe):
+ gpus = gpus.split("-")
+ os.makedirs("%s/logs/%s" % (now_dir, exp_dir), exist_ok=True)
+ f = open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "w")
+ f.close()
+ if if_f0:
+ if f0method != "rmvpe_gpu":
+ cmd = (
+ '"%s" infer/modules/train/extract/extract_f0_print.py "%s/logs/%s" %s %s'
+ % (
+ config.python_cmd,
+ now_dir,
+ exp_dir,
+ n_p,
+ f0method,
+ )
+ )
+ logger.info("Execute: " + cmd)
+ p = Popen(
+ cmd, shell=True, cwd=now_dir
+ ) # , stdin=PIPE, stdout=PIPE,stderr=PIPE
+ # 煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
+ done = [False]
+ threading.Thread(
+ target=if_done,
+ args=(
+ done,
+ p,
+ ),
+ ).start()
+ else:
+ if gpus_rmvpe != "-":
+ gpus_rmvpe = gpus_rmvpe.split("-")
+ leng = len(gpus_rmvpe)
+ ps = []
+ for idx, n_g in enumerate(gpus_rmvpe):
+ cmd = (
+ '"%s" infer/modules/train/extract/extract_f0_rmvpe.py %s %s %s "%s/logs/%s" %s '
+ % (
+ config.python_cmd,
+ leng,
+ idx,
+ n_g,
+ now_dir,
+ exp_dir,
+ config.is_half,
+ )
+ )
+ logger.info("Execute: " + cmd)
+ p = Popen(
+ cmd, shell=True, cwd=now_dir
+ ) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
+ ps.append(p)
+ # 煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
+ done = [False]
+ threading.Thread(
+ target=if_done_multi, #
+ args=(
+ done,
+ ps,
+ ),
+ ).start()
+ else:
+ cmd = (
+ config.python_cmd
+ + ' infer/modules/train/extract/extract_f0_rmvpe_dml.py "%s/logs/%s" '
+ % (
+ now_dir,
+ exp_dir,
+ )
+ )
+ logger.info("Execute: " + cmd)
+ p = Popen(
+ cmd, shell=True, cwd=now_dir
+ ) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
+ p.wait()
+ done = [True]
+ while 1:
+ with open(
+ "%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r"
+ ) as f:
+ yield (f.read())
+ sleep(1)
+ if done[0]:
+ break
+ with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
+ log = f.read()
+ logger.info(log)
+ yield log
+ # 对不同part分别开多进程
+ """
+ n_part=int(sys.argv[1])
+ i_part=int(sys.argv[2])
+ i_gpu=sys.argv[3]
+ exp_dir=sys.argv[4]
+ os.environ["CUDA_VISIBLE_DEVICES"]=str(i_gpu)
+ """
+ leng = len(gpus)
+ ps = []
+ for idx, n_g in enumerate(gpus):
+ cmd = (
+ '"%s" infer/modules/train/extract_feature_print.py %s %s %s %s "%s/logs/%s" %s %s'
+ % (
+ config.python_cmd,
+ config.device,
+ leng,
+ idx,
+ n_g,
+ now_dir,
+ exp_dir,
+ version19,
+ config.is_half,
+ )
+ )
+ logger.info("Execute: " + cmd)
+ p = Popen(
+ cmd, shell=True, cwd=now_dir
+ ) # , shell=True, stdin=PIPE, stdout=PIPE, stderr=PIPE, cwd=now_dir
+ ps.append(p)
+ # 煞笔gr, popen read都非得全跑完了再一次性读取, 不用gr就正常读一句输出一句;只能额外弄出一个文本流定时读
+ done = [False]
+ threading.Thread(
+ target=if_done_multi,
+ args=(
+ done,
+ ps,
+ ),
+ ).start()
+ while 1:
+ with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
+ yield (f.read())
+ sleep(1)
+ if done[0]:
+ break
+ with open("%s/logs/%s/extract_f0_feature.log" % (now_dir, exp_dir), "r") as f:
+ log = f.read()
+ logger.info(log)
+ yield log
+
+
+def get_pretrained_models(path_str, f0_str, sr2):
+ if_pretrained_generator_exist = os.access(
+ "assets/pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2), os.F_OK
+ )
+ if_pretrained_discriminator_exist = os.access(
+ "assets/pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2), os.F_OK
+ )
+ if not if_pretrained_generator_exist:
+ logger.warning(
+ "assets/pretrained%s/%sG%s.pth not exist, will not use pretrained model",
+ path_str,
+ f0_str,
+ sr2,
+ )
+ if not if_pretrained_discriminator_exist:
+ logger.warning(
+ "assets/pretrained%s/%sD%s.pth not exist, will not use pretrained model",
+ path_str,
+ f0_str,
+ sr2,
+ )
+ return (
+ (
+ "assets/pretrained%s/%sG%s.pth" % (path_str, f0_str, sr2)
+ if if_pretrained_generator_exist
+ else ""
+ ),
+ (
+ "assets/pretrained%s/%sD%s.pth" % (path_str, f0_str, sr2)
+ if if_pretrained_discriminator_exist
+ else ""
+ ),
+ )
+
+
+def change_sr2(sr2, if_f0_3, version19):
+ path_str = "" if version19 == "v1" else "_v2"
+ f0_str = "f0" if if_f0_3 else ""
+ return get_pretrained_models(path_str, f0_str, sr2)
+
+
+def change_version19(sr2, if_f0_3, version19):
+ path_str = "" if version19 == "v1" else "_v2"
+ if sr2 == "32k" and version19 == "v1":
+ sr2 = "40k"
+ to_return_sr2 = (
+ {"choices": ["40k", "48k"], "__type__": "update", "value": sr2}
+ if version19 == "v1"
+ else {"choices": ["40k", "48k", "32k"], "__type__": "update", "value": sr2}
+ )
+ f0_str = "f0" if if_f0_3 else ""
+ return (
+ *get_pretrained_models(path_str, f0_str, sr2),
+ to_return_sr2,
+ )
+
+
+def change_f0(if_f0_3, sr2, version19): # f0method8,pretrained_G14,pretrained_D15
+ path_str = "" if version19 == "v1" else "_v2"
+ return (
+ {"visible": if_f0_3, "__type__": "update"},
+ {"visible": if_f0_3, "__type__": "update"},
+ *get_pretrained_models(path_str, "f0" if if_f0_3 == True else "", sr2),
+ )
+
+
+# but3.click(click_train,[exp_dir1,sr2,if_f0_3,save_epoch10,total_epoch11,batch_size12,if_save_latest13,pretrained_G14,pretrained_D15,gpus16])
+def click_train(
+ exp_dir1,
+ sr2,
+ if_f0_3,
+ spk_id5,
+ save_epoch10,
+ total_epoch11,
+ batch_size12,
+ if_save_latest13,
+ pretrained_G14,
+ pretrained_D15,
+ gpus16,
+ if_cache_gpu17,
+ if_save_every_weights18,
+ version19,
+ author,
+):
+ # 生成filelist
+ exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
+ os.makedirs(exp_dir, exist_ok=True)
+ gt_wavs_dir = "%s/0_gt_wavs" % (exp_dir)
+ feature_dir = (
+ "%s/3_feature256" % (exp_dir)
+ if version19 == "v1"
+ else "%s/3_feature768" % (exp_dir)
+ )
+ if if_f0_3:
+ f0_dir = "%s/2a_f0" % (exp_dir)
+ f0nsf_dir = "%s/2b-f0nsf" % (exp_dir)
+ names = (
+ set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)])
+ & set([name.split(".")[0] for name in os.listdir(feature_dir)])
+ & set([name.split(".")[0] for name in os.listdir(f0_dir)])
+ & set([name.split(".")[0] for name in os.listdir(f0nsf_dir)])
+ )
+ else:
+ names = set([name.split(".")[0] for name in os.listdir(gt_wavs_dir)]) & set(
+ [name.split(".")[0] for name in os.listdir(feature_dir)]
+ )
+ opt = []
+ for name in names:
+ if if_f0_3:
+ opt.append(
+ "%s/%s.wav|%s/%s.npy|%s/%s.wav.npy|%s/%s.wav.npy|%s"
+ % (
+ gt_wavs_dir.replace("\\", "\\\\"),
+ name,
+ feature_dir.replace("\\", "\\\\"),
+ name,
+ f0_dir.replace("\\", "\\\\"),
+ name,
+ f0nsf_dir.replace("\\", "\\\\"),
+ name,
+ spk_id5,
+ )
+ )
+ else:
+ opt.append(
+ "%s/%s.wav|%s/%s.npy|%s"
+ % (
+ gt_wavs_dir.replace("\\", "\\\\"),
+ name,
+ feature_dir.replace("\\", "\\\\"),
+ name,
+ spk_id5,
+ )
+ )
+ fea_dim = 256 if version19 == "v1" else 768
+ if if_f0_3:
+ for _ in range(2):
+ opt.append(
+ "%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s/logs/mute/2a_f0/mute.wav.npy|%s/logs/mute/2b-f0nsf/mute.wav.npy|%s"
+ % (now_dir, sr2, now_dir, fea_dim, now_dir, now_dir, spk_id5)
+ )
+ else:
+ for _ in range(2):
+ opt.append(
+ "%s/logs/mute/0_gt_wavs/mute%s.wav|%s/logs/mute/3_feature%s/mute.npy|%s"
+ % (now_dir, sr2, now_dir, fea_dim, spk_id5)
+ )
+ shuffle(opt)
+ with open("%s/filelist.txt" % exp_dir, "w") as f:
+ f.write("\n".join(opt))
+ logger.debug("Write filelist done")
+ logger.info("Use gpus: %s", str(gpus16))
+ if pretrained_G14 == "":
+ logger.info("No pretrained Generator")
+ if pretrained_D15 == "":
+ logger.info("No pretrained Discriminator")
+ if version19 == "v1" or sr2 == "40k":
+ config_path = "v1/%s.json" % sr2
+ else:
+ config_path = "v2/%s.json" % sr2
+ config_save_path = os.path.join(exp_dir, "config.json")
+ if not pathlib.Path(config_save_path).exists():
+ with open(config_save_path, "w", encoding="utf-8") as f:
+ json.dump(
+ config.json_config[config_path],
+ f,
+ ensure_ascii=False,
+ indent=4,
+ sort_keys=True,
+ )
+ f.write("\n")
+ cmd = (
+ '"%s" infer/modules/train/train.py -e "%s" -sr %s -f0 %s -bs %s -te %s -se %s %s %s -l %s -c %s -sw %s -v %s -a "%s"'
+ % (
+ config.python_cmd,
+ exp_dir1,
+ sr2,
+ 1 if if_f0_3 else 0,
+ batch_size12,
+ total_epoch11,
+ save_epoch10,
+ '-pg "%s"' % pretrained_G14 if pretrained_G14 != "" else "",
+ '-pd "%s"' % pretrained_D15 if pretrained_D15 != "" else "",
+ 1 if if_save_latest13 == i18n("Yes") else 0,
+ 1 if if_cache_gpu17 == i18n("Yes") else 0,
+ 1 if if_save_every_weights18 == i18n("Yes") else 0,
+ version19,
+ author,
+ )
+ )
+ if gpus16:
+ cmd += ' -g "%s"' % (gpus16)
+
+ logger.info("Execute: " + cmd)
+ p = Popen(cmd, shell=True, cwd=now_dir)
+ p.wait()
+ return "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder."
+
+
+# but4.click(train_index, [exp_dir1], info3)
+def train_index(exp_dir1, version19):
+ # exp_dir = "%s/logs/%s" % (now_dir, exp_dir1)
+ exp_dir = "logs/%s" % (exp_dir1)
+ os.makedirs(exp_dir, exist_ok=True)
+ feature_dir = (
+ "%s/3_feature256" % (exp_dir)
+ if version19 == "v1"
+ else "%s/3_feature768" % (exp_dir)
+ )
+ if not os.path.exists(feature_dir):
+ return "请先进行特征提取!"
+ listdir_res = list(os.listdir(feature_dir))
+ if len(listdir_res) == 0:
+ return "请先进行特征提取！"
+ infos = []
+ npys = []
+ for name in sorted(listdir_res):
+ phone = np.load("%s/%s" % (feature_dir, name))
+ npys.append(phone)
+ big_npy = np.concatenate(npys, 0)
+ big_npy_idx = np.arange(big_npy.shape[0])
+ np.random.shuffle(big_npy_idx)
+ big_npy = big_npy[big_npy_idx]
+ if big_npy.shape[0] > 2e5:
+ infos.append("Trying doing kmeans %s shape to 10k centers." % big_npy.shape[0])
+ yield "\n".join(infos)
+ try:
+ big_npy = (
+ MiniBatchKMeans(
+ n_clusters=10000,
+ verbose=True,
+ batch_size=256 * config.n_cpu,
+ compute_labels=False,
+ init="random",
+ )
+ .fit(big_npy)
+ .cluster_centers_
+ )
+ except:
+ info = traceback.format_exc()
+ logger.info(info)
+ infos.append(info)
+ yield "\n".join(infos)
+
+ np.save("%s/total_fea.npy" % exp_dir, big_npy)
+ n_ivf = min(int(16 * np.sqrt(big_npy.shape[0])), big_npy.shape[0] // 39)
+ infos.append("%s,%s" % (big_npy.shape, n_ivf))
+ yield "\n".join(infos)
+ index = faiss.index_factory(256 if version19 == "v1" else 768, "IVF%s,Flat" % n_ivf)
+ # index = faiss.index_factory(256if version19=="v1"else 768, "IVF%s,PQ128x4fs,RFlat"%n_ivf)
+ infos.append("training")
+ yield "\n".join(infos)
+ index_ivf = faiss.extract_index_ivf(index) #
+ index_ivf.nprobe = 1
+ index.train(big_npy)
+ faiss.write_index(
+ index,
+ "%s/trained_IVF%s_Flat_nprobe_%s_%s_%s.index"
+ % (exp_dir, n_ivf, index_ivf.nprobe, exp_dir1, version19),
+ )
+ infos.append("adding")
+ yield "\n".join(infos)
+ batch_size_add = 8192
+ for i in range(0, big_npy.shape[0], batch_size_add):
+ index.add(big_npy[i : i + batch_size_add])
+ index_save_path = "%s/added_IVF%s_Flat_nprobe_%s_%s_%s.index" % (
+ exp_dir,
+ n_ivf,
+ index_ivf.nprobe,
+ exp_dir1,
+ version19,
+ )
+ faiss.write_index(index, index_save_path)
+ infos.append(i18n("Successfully built index into") + " " + index_save_path)
+ link_target = "%s/%s_IVF%s_Flat_nprobe_%s_%s_%s.index" % (
+ outside_index_root,
+ exp_dir1,
+ n_ivf,
+ index_ivf.nprobe,
+ exp_dir1,
+ version19,
+ )
+ try:
+ link = os.link if platform.system() == "Windows" else os.symlink
+ link(index_save_path, link_target)
+ infos.append(i18n("Link index to outside folder") + " " + link_target)
+ except:
+ infos.append(
+ i18n("Link index to outside folder")
+ + " "
+ + link_target
+ + " "
+ + i18n("Fail")
+ )
+
+ # faiss.write_index(index, '%s/added_IVF%s_Flat_FastScan_%s.index'%(exp_dir,n_ivf,version19))
+ # infos.append("成功构建索引，added_IVF%s_Flat_FastScan_%s.index"%(n_ivf,version19))
+ yield "\n".join(infos)
+
+
+# but5.click(train1key, [exp_dir1, sr2, if_f0_3, trainset_dir4, spk_id5, gpus6, np7, f0method8, save_epoch10, total_epoch11, batch_size12, if_save_latest13, pretrained_G14, pretrained_D15, gpus16, if_cache_gpu17], info3)
+def train1key(
+ exp_dir1,
+ sr2,
+ if_f0_3,
+ trainset_dir4,
+ spk_id5,
+ np7,
+ f0method8,
+ save_epoch10,
+ total_epoch11,
+ batch_size12,
+ if_save_latest13,
+ pretrained_G14,
+ pretrained_D15,
+ gpus16,
+ if_cache_gpu17,
+ if_save_every_weights18,
+ version19,
+ gpus_rmvpe,
+ author,
+):
+ infos = []
+
+ def get_info_str(strr):
+ infos.append(strr)
+ return "\n".join(infos)
+
+ # step1:Process data
+ yield get_info_str(i18n("Step 1: Processing data"))
+ [get_info_str(_) for _ in preprocess_dataset(trainset_dir4, exp_dir1, sr2, np7)]
+
+ # step2a:提取音高
+ yield get_info_str(i18n("step2:Pitch extraction & feature extraction"))
+ [
+ get_info_str(_)
+ for _ in extract_f0_feature(
+ gpus16, np7, f0method8, if_f0_3, exp_dir1, version19, gpus_rmvpe
+ )
+ ]
+
+ # step3a:Train model
+ yield get_info_str(i18n("Step 3a: Model training started"))
+ click_train(
+ exp_dir1,
+ sr2,
+ if_f0_3,
+ spk_id5,
+ save_epoch10,
+ total_epoch11,
+ batch_size12,
+ if_save_latest13,
+ pretrained_G14,
+ pretrained_D15,
+ gpus16,
+ if_cache_gpu17,
+ if_save_every_weights18,
+ version19,
+ author,
+ )
+ yield get_info_str(
+ i18n(
+ "Training complete. You can check the training logs in the console or the 'train.log' file under the experiment folder."
+ )
+ )
+
+ # step3b:训练索引
+ [get_info_str(_) for _ in train_index(exp_dir1, version19)]
+ yield get_info_str(i18n("All processes have been completed!"))
+
+
+# ckpt_path2.change(change_info_,[ckpt_path2],[sr__,if_f0__])
+def change_info_(ckpt_path):
+ if not os.path.exists(ckpt_path.replace(os.path.basename(ckpt_path), "train.log")):
+ return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
+ try:
+ with open(
+ ckpt_path.replace(os.path.basename(ckpt_path), "train.log"), "r"
+ ) as f:
+ info = eval(f.read().strip("\n").split("\n")[0].split("\t")[-1])
+ sr, f0 = info["sample_rate"], info["if_f0"]
+ version = "v2" if ("version" in info and info["version"] == "v2") else "v1"
+ return sr, str(f0), version
+ except:
+ traceback.print_exc()
+ return {"__type__": "update"}, {"__type__": "update"}, {"__type__": "update"}
+
+
+F0GPUVisible = config.dml == False
+
+
+def change_f0_method(f0method8):
+ if f0method8 == "rmvpe_gpu":
+ visible = F0GPUVisible
+ else:
+ visible = False
+ return {"visible": visible, "__type__": "update"}
+
+
+with gr.Blocks(title="RVC WebUI") as app:
+ gr.Markdown("## RVC WebUI")
+ gr.Markdown(
+ value=i18n(
+ "This software is open source under the MIT license. The author does not have any control over the software. Users who use the software and distribute the sounds exported by the software are solely responsible.
If you do not agree with this clause, you cannot use or reference any codes and files within the software package. See the root directory Agreement-LICENSE.txt for details."
+ )
+ )
+ with gr.Tabs():
+ with gr.TabItem(i18n("Model Inference")):
+ with gr.Row():
+ sid0 = gr.Dropdown(
+ label=i18n("Inferencing voice"), choices=sorted(names)
+ )
+ with gr.Column():
+ refresh_button = gr.Button(
+ i18n("Refresh voice list and index path"), variant="primary"
+ )
+ clean_button = gr.Button(
+ i18n("Unload model to save GPU memory"), variant="primary"
+ )
+ spk_item = gr.Slider(
+ minimum=0,
+ maximum=2333,
+ step=1,
+ label=i18n("Select Speaker/Singer ID"),
+ value=0,
+ visible=False,
+ interactive=True,
+ )
+ clean_button.click(
+ fn=clean, inputs=[], outputs=[sid0], api_name="infer_clean"
+ )
+ modelinfo = gr.Textbox(label=i18n("Model info"), max_lines=8)
+ with gr.TabItem(i18n("Single inference")):
+ with gr.Row():
+ with gr.Column():
+ vc_transform0 = gr.Number(
+ label=i18n(
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)"
+ ),
+ value=0,
+ )
+ input_audio0 = gr.Audio(
+ label=i18n("The audio file to be processed"),
+ type="filepath",
+ )
+ file_index2 = gr.Dropdown(
+ label=i18n(
+ "Auto-detect index path and select from the dropdown"
+ ),
+ choices=sorted(index_paths),
+ interactive=True,
+ )
+ file_index1 = gr.File(
+ label=i18n(
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown"
+ ),
+ )
+ with gr.Column():
+ f0method0 = gr.Radio(
+ label=i18n(
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement"
+ ),
+ choices=(
+ ["pm", "dio", "harvest", "crepe", "rmvpe", "fcpe"]
+ ),
+ value="rmvpe",
+ interactive=True,
+ )
+ resample_sr0 = gr.Slider(
+ minimum=0,
+ maximum=48000,
+ label=i18n(
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling"
+ ),
+ value=0,
+ step=1,
+ interactive=True,
+ )
+ rms_mix_rate0 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n(
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume"
+ ),
+ value=0.25,
+ interactive=True,
+ )
+ protect0 = gr.Slider(
+ minimum=0,
+ maximum=0.5,
+ label=i18n(
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy"
+ ),
+ value=0.33,
+ step=0.01,
+ interactive=True,
+ )
+ filter_radius0 = gr.Slider(
+ minimum=0,
+ maximum=7,
+ label=i18n(
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness."
+ ),
+ value=3,
+ step=1,
+ interactive=True,
+ )
+ index_rate1 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n("Feature searching ratio"),
+ value=0.75,
+ interactive=True,
+ )
+ f0_file = gr.File(
+ label=i18n(
+ "F0 curve file (optional). One pitch per line. Replaces the default F0 and pitch modulation"
+ ),
+ visible=False,
+ )
+ but0 = gr.Button(i18n("Convert"), variant="primary")
+ vc_output2 = gr.Audio(
+ label=i18n(
+ "Export audio (click on the three dots in the lower right corner to download)"
+ )
+ )
+
+ refresh_button.click(
+ fn=change_choices,
+ inputs=[],
+ outputs=[sid0, file_index2],
+ api_name="infer_refresh",
+ )
+
+ vc_output1 = gr.Textbox(label=i18n("Output information"))
+
+ but0.click(
+ vc.vc_single,
+ [
+ spk_item,
+ input_audio0,
+ vc_transform0,
+ f0_file,
+ f0method0,
+ file_index1,
+ file_index2,
+ # file_big_npy1,
+ index_rate1,
+ filter_radius0,
+ resample_sr0,
+ rms_mix_rate0,
+ protect0,
+ ],
+ [vc_output1, vc_output2],
+ api_name="infer_convert",
+ )
+ with gr.TabItem(i18n("Batch inference")):
+ gr.Markdown(
+ value=i18n(
+ "Batch conversion. Enter the folder containing the audio files to be converted or upload multiple audio files. The converted audio will be output in the specified folder (default: 'opt')."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ vc_transform1 = gr.Number(
+ label=i18n(
+ "Transpose (integer, number of semitones, raise by an octave: 12, lower by an octave: -12)"
+ ),
+ value=0,
+ )
+ dir_input = gr.Textbox(
+ label=i18n(
+ "Enter the path of the audio folder to be processed (copy it from the address bar of the file manager)"
+ ),
+ placeholder="C:\\Users\\Desktop\\input_vocal_dir",
+ )
+ inputs = gr.File(
+ file_count="multiple",
+ label=i18n(
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored."
+ ),
+ )
+ opt_input = gr.Textbox(
+ label=i18n("Specify output folder"), value="opt"
+ )
+ file_index4 = gr.Dropdown(
+ label=i18n(
+ "Auto-detect index path and select from the dropdown"
+ ),
+ choices=sorted(index_paths),
+ interactive=True,
+ )
+ file_index3 = gr.File(
+ label=i18n(
+ "Path to the feature index file. Leave blank to use the selected result from the dropdown"
+ ),
+ )
+
+ refresh_button.click(
+ fn=lambda: change_choices()[1],
+ inputs=[],
+ outputs=file_index4,
+ api_name="infer_refresh_batch",
+ )
+ # file_big_npy2 = gr.Textbox(
+ # label=i18n("特征文件路径"),
+ # value="E:\\codes\\py39\\vits_vc_gpu_train\\logs\\mi-test-1key\\total_fea.npy",
+ # interactive=True,
+ # )
+
+ with gr.Column():
+ f0method1 = gr.Radio(
+ label=i18n(
+ "Select the pitch extraction algorithm ('pm': faster extraction but lower-quality speech; 'harvest': better bass but extremely slow; 'crepe': better quality but GPU intensive), 'rmvpe': best quality, and little GPU requirement"
+ ),
+ choices=(
+ ["pm", "harvest", "crepe", "rmvpe"]
+ if config.dml == False
+ else ["pm", "harvest", "rmvpe"]
+ ),
+ value="rmvpe",
+ interactive=True,
+ )
+ resample_sr1 = gr.Slider(
+ minimum=0,
+ maximum=48000,
+ label=i18n(
+ "Resample the output audio in post-processing to the final sample rate. Set to 0 for no resampling"
+ ),
+ value=0,
+ step=1,
+ interactive=True,
+ )
+ rms_mix_rate1 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n(
+ "Adjust the volume envelope scaling. Closer to 0, the more it mimicks the volume of the original vocals. Can help mask noise and make volume sound more natural when set relatively low. Closer to 1 will be more of a consistently loud volume"
+ ),
+ value=1,
+ interactive=True,
+ )
+ protect1 = gr.Slider(
+ minimum=0,
+ maximum=0.5,
+ label=i18n(
+ "Protect voiceless consonants and breath sounds to prevent artifacts such as tearing in electronic music. Set to 0.5 to disable. Decrease the value to increase protection, but it may reduce indexing accuracy"
+ ),
+ value=0.33,
+ step=0.01,
+ interactive=True,
+ )
+ filter_radius1 = gr.Slider(
+ minimum=0,
+ maximum=7,
+ label=i18n(
+ "If >=3: apply median filtering to the harvested pitch results. The value represents the filter radius and can reduce breathiness."
+ ),
+ value=3,
+ step=1,
+ interactive=True,
+ )
+ index_rate2 = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n("Feature searching ratio"),
+ value=1,
+ interactive=True,
+ )
+ format1 = gr.Radio(
+ label=i18n("Export file format"),
+ choices=["wav", "flac", "mp3", "m4a"],
+ value="wav",
+ interactive=True,
+ )
+ but1 = gr.Button(i18n("Convert"), variant="primary")
+ vc_output3 = gr.Textbox(label=i18n("Output information"))
+
+ but1.click(
+ vc.vc_multi,
+ [
+ spk_item,
+ dir_input,
+ opt_input,
+ inputs,
+ vc_transform1,
+ f0method1,
+ file_index3,
+ file_index4,
+ # file_big_npy2,
+ index_rate2,
+ filter_radius1,
+ resample_sr1,
+ rms_mix_rate1,
+ protect1,
+ format1,
+ ],
+ [vc_output3],
+ api_name="infer_convert_batch",
+ )
+ sid0.change(
+ fn=vc.get_vc,
+ inputs=[sid0, protect0, protect1, file_index2, file_index4],
+ outputs=[
+ spk_item,
+ protect0,
+ protect1,
+ file_index2,
+ file_index4,
+ modelinfo,
+ ],
+ api_name="infer_change_voice",
+ )
+ with gr.TabItem(
+ i18n("Vocals/Accompaniment Separation & Reverberation Removal")
+ ):
+ gr.Markdown(
+ value=i18n(
+ "Batch processing for vocal accompaniment separation using the UVR5 model.
Example of a valid folder path format: D:\\path\\to\\input\\folder (copy it from the file manager address bar).
The model is divided into three categories:
1. Preserve vocals: Choose this option for audio without harmonies. It preserves vocals better than HP5. It includes two built-in models: HP2 and HP3. HP3 may slightly leak accompaniment but preserves vocals slightly better than HP2.
2. Preserve main vocals only: Choose this option for audio with harmonies. It may weaken the main vocals. It includes one built-in model: HP5.
3. De-reverb and de-delay models (by FoxJoy):
(1) MDX-Net: The best choice for stereo reverb removal but cannot remove mono reverb;
(234) DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverb and can remove mono reverb, but not very effectively for heavily reverberated high-frequency content.
De-reverb/de-delay notes:
1. The processing time for the DeEcho-DeReverb model is approximately twice as long as the other two DeEcho models.
2. The MDX-Net-Dereverb model is quite slow.
3. The recommended cleanest configuration is to apply MDX-Net first and then DeEcho-Aggressive."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ dir_wav_input = gr.Textbox(
+ label=i18n(
+ "Enter the path of the audio folder to be processed"
+ ),
+ placeholder="C:\\Users\\Desktop\\todo-songs",
+ )
+ wav_inputs = gr.File(
+ file_count="multiple",
+ label=i18n(
+ "Multiple audio files can also be imported. If a folder path exists, this input is ignored."
+ ),
+ )
+ with gr.Column():
+ model_choose = gr.Dropdown(label=i18n("Model"), choices=uvr5_names)
+ agg = gr.Slider(
+ minimum=0,
+ maximum=20,
+ step=1,
+ label="人声提取激进程度",
+ value=10,
+ interactive=True,
+ visible=False, # 先不开放调整
+ )
+ opt_vocal_root = gr.Textbox(
+ label=i18n("Specify the output folder for vocals"),
+ value="opt",
+ )
+ opt_ins_root = gr.Textbox(
+ label=i18n("Specify the output folder for accompaniment"),
+ value="opt",
+ )
+ format0 = gr.Radio(
+ label=i18n("Export file format"),
+ choices=["wav", "flac", "mp3", "m4a"],
+ value="flac",
+ interactive=True,
+ )
+ but2 = gr.Button(i18n("Convert"), variant="primary")
+ vc_output4 = gr.Textbox(label=i18n("Output information"))
+ but2.click(
+ uvr,
+ [
+ model_choose,
+ dir_wav_input,
+ opt_vocal_root,
+ wav_inputs,
+ opt_ins_root,
+ agg,
+ format0,
+ ],
+ [vc_output4],
+ api_name="uvr_convert",
+ )
+ with gr.TabItem(i18n("Train")):
+ gr.Markdown(
+ value=i18n(
+ "### Step 1. Fill in the experimental configuration.\nExperimental data is stored in the 'logs' folder, with each experiment having a separate folder. Manually enter the experiment name path, which contains the experimental configuration, logs, and trained model files."
+ )
+ )
+ with gr.Row():
+ exp_dir1 = gr.Textbox(
+ label=i18n("Enter the experiment name"), value="mi-test"
+ )
+ author = gr.Textbox(label=i18n("Model Author (Nullable)"))
+ np7 = gr.Slider(
+ minimum=0,
+ maximum=config.n_cpu,
+ step=1,
+ label=i18n(
+ "Number of CPU processes used for pitch extraction and data processing"
+ ),
+ value=int(np.ceil(config.n_cpu / 1.5)),
+ interactive=True,
+ )
+ with gr.Row():
+ sr2 = gr.Radio(
+ label=i18n("Target sample rate"),
+ choices=["40k", "48k"],
+ value="40k",
+ interactive=True,
+ )
+ if_f0_3 = gr.Radio(
+ label=i18n(
+ "Whether the model has pitch guidance (required for singing, optional for speech)"
+ ),
+ choices=[i18n("Yes"), i18n("No")],
+ value=i18n("Yes"),
+ interactive=True,
+ )
+ version19 = gr.Radio(
+ label=i18n("Version"),
+ choices=["v1", "v2"],
+ value="v2",
+ interactive=True,
+ visible=True,
+ )
+ gr.Markdown(
+ value=i18n(
+ "### Step 2. Audio processing. \n#### 1. Slicing.\nAutomatically traverse all files in the training folder that can be decoded into audio and perform slice normalization. Generates 2 wav folders in the experiment directory. Currently, only single-singer/speaker training is supported."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ trainset_dir4 = gr.Textbox(
+ label=i18n("Enter the path of the training folder"),
+ )
+ spk_id5 = gr.Slider(
+ minimum=0,
+ maximum=4,
+ step=1,
+ label=i18n("Please specify the speaker/singer ID"),
+ value=0,
+ interactive=True,
+ )
+ but1 = gr.Button(i18n("Process data"), variant="primary")
+ with gr.Column():
+ info1 = gr.Textbox(label=i18n("Output information"), value="")
+ but1.click(
+ preprocess_dataset,
+ [trainset_dir4, exp_dir1, sr2, np7],
+ [info1],
+ api_name="train_preprocess",
+ )
+ gr.Markdown(
+ value=i18n(
+ "#### 2. Feature extraction.\nUse CPU to extract pitch (if the model has pitch), use GPU to extract features (select GPU index)."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ gpu_info9 = gr.Textbox(
+ label=i18n("GPU Information"),
+ value=gpu_info,
+ visible=F0GPUVisible,
+ )
+ gpus6 = gr.Textbox(
+ label=i18n(
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2"
+ ),
+ value=gpus,
+ interactive=True,
+ visible=F0GPUVisible,
+ )
+ gpus_rmvpe = gr.Textbox(
+ label=i18n(
+ "Enter the GPU index(es) separated by '-', e.g., 0-0-1 to use 2 processes in GPU0 and 1 process in GPU1"
+ ),
+ value="%s-%s" % (gpus, gpus),
+ interactive=True,
+ visible=F0GPUVisible,
+ )
+ f0method8 = gr.Radio(
+ label=i18n(
+ "Select the pitch extraction algorithm: when extracting singing, you can use 'pm' to speed up. For high-quality speech with fast performance, but worse CPU usage, you can use 'dio'. 'harvest' results in better quality but is slower. 'rmvpe' has the best results and consumes less CPU/GPU"
+ ),
+ choices=["pm", "harvest", "dio", "rmvpe", "rmvpe_gpu"],
+ value="rmvpe_gpu",
+ interactive=True,
+ )
+ with gr.Column():
+ but2 = gr.Button(i18n("Feature extraction"), variant="primary")
+ info2 = gr.Textbox(label=i18n("Output information"), value="")
+ f0method8.change(
+ fn=change_f0_method,
+ inputs=[f0method8],
+ outputs=[gpus_rmvpe],
+ )
+ but2.click(
+ extract_f0_feature,
+ [
+ gpus6,
+ np7,
+ f0method8,
+ if_f0_3,
+ exp_dir1,
+ version19,
+ gpus_rmvpe,
+ ],
+ [info2],
+ api_name="train_extract_f0_feature",
+ )
+ gr.Markdown(
+ value=i18n(
+ "### Step 3. Start training.\nFill in the training settings and start training the model and index."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ save_epoch10 = gr.Slider(
+ minimum=1,
+ maximum=50,
+ step=1,
+ label=i18n("Save frequency (save_every_epoch)"),
+ value=5,
+ interactive=True,
+ )
+ total_epoch11 = gr.Slider(
+ minimum=2,
+ maximum=1000,
+ step=1,
+ label=i18n("Total training epochs (total_epoch)"),
+ value=20,
+ interactive=True,
+ )
+ batch_size12 = gr.Slider(
+ minimum=1,
+ maximum=40,
+ step=1,
+ label=i18n("Batch size per GPU"),
+ value=default_batch_size,
+ interactive=True,
+ )
+ if_save_latest13 = gr.Radio(
+ label=i18n(
+ "Save only the latest '.ckpt' file to save disk space"
+ ),
+ choices=[i18n("Yes"), i18n("No")],
+ value=i18n("No"),
+ interactive=True,
+ )
+ if_cache_gpu17 = gr.Radio(
+ label=i18n(
+ "Cache all training sets to GPU memory. Caching small datasets (less than 10 minutes) can speed up training, but caching large datasets will consume a lot of GPU memory and may not provide much speed improvement"
+ ),
+ choices=[i18n("Yes"), i18n("No")],
+ value=i18n("No"),
+ interactive=True,
+ )
+ if_save_every_weights18 = gr.Radio(
+ label=i18n(
+ "Save a small final model to the 'weights' folder at each save point"
+ ),
+ choices=[i18n("Yes"), i18n("No")],
+ value=i18n("No"),
+ interactive=True,
+ )
+ with gr.Column():
+ pretrained_G14 = gr.Textbox(
+ label=i18n("Load pre-trained base model G path"),
+ value="assets/pretrained_v2/f0G40k.pth",
+ interactive=True,
+ )
+ pretrained_D15 = gr.Textbox(
+ label=i18n("Load pre-trained base model D path"),
+ value="assets/pretrained_v2/f0D40k.pth",
+ interactive=True,
+ )
+ gpus16 = gr.Textbox(
+ label=i18n(
+ "Enter the GPU index(es) separated by '-', e.g., 0-1-2 to use GPU 0, 1, and 2"
+ ),
+ value=gpus,
+ interactive=True,
+ )
+ sr2.change(
+ change_sr2,
+ [sr2, if_f0_3, version19],
+ [pretrained_G14, pretrained_D15],
+ )
+ version19.change(
+ change_version19,
+ [sr2, if_f0_3, version19],
+ [pretrained_G14, pretrained_D15, sr2],
+ )
+ if_f0_3.change(
+ change_f0,
+ [if_f0_3, sr2, version19],
+ [f0method8, gpus_rmvpe, pretrained_G14, pretrained_D15],
+ )
+
+ but3 = gr.Button(i18n("Train model"), variant="primary")
+ but4 = gr.Button(i18n("Train feature index"), variant="primary")
+ but5 = gr.Button(i18n("One-click training"), variant="primary")
+ with gr.Row():
+ info3 = gr.Textbox(label=i18n("Output information"), value="")
+ but3.click(
+ click_train,
+ [
+ exp_dir1,
+ sr2,
+ if_f0_3,
+ spk_id5,
+ save_epoch10,
+ total_epoch11,
+ batch_size12,
+ if_save_latest13,
+ pretrained_G14,
+ pretrained_D15,
+ gpus16,
+ if_cache_gpu17,
+ if_save_every_weights18,
+ version19,
+ author,
+ ],
+ info3,
+ api_name="train_start",
+ )
+ but4.click(train_index, [exp_dir1, version19], info3)
+ but5.click(
+ train1key,
+ [
+ exp_dir1,
+ sr2,
+ if_f0_3,
+ trainset_dir4,
+ spk_id5,
+ np7,
+ f0method8,
+ save_epoch10,
+ total_epoch11,
+ batch_size12,
+ if_save_latest13,
+ pretrained_G14,
+ pretrained_D15,
+ gpus16,
+ if_cache_gpu17,
+ if_save_every_weights18,
+ version19,
+ gpus_rmvpe,
+ author,
+ ],
+ info3,
+ api_name="train_start_all",
+ )
+
+ with gr.TabItem(i18n("ckpt Processing")):
+ gr.Markdown(
+ value=i18n(
+ "### Model comparison\n> You can get model ID (long) from `View model information` below.\n\nCalculate a similarity between two models."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ id_a = gr.Textbox(label=i18n("ID of model A (long)"), value="")
+ id_b = gr.Textbox(label=i18n("ID of model B (long)"), value="")
+ with gr.Column():
+ butmodelcmp = gr.Button(i18n("Calculate"), variant="primary")
+ infomodelcmp = gr.Textbox(
+ label=i18n("Similarity (from 0 to 1)"),
+ value="",
+ max_lines=1,
+ )
+ butmodelcmp.click(
+ hash_similarity,
+ [
+ id_a,
+ id_b,
+ ],
+ infomodelcmp,
+ api_name="ckpt_merge",
+ )
+
+ gr.Markdown(
+ value=i18n("### Model fusion\nCan be used to test timbre fusion.")
+ )
+ with gr.Row():
+ with gr.Column():
+ ckpt_a = gr.Textbox(
+ label=i18n("Path to Model A"), value="", interactive=True
+ )
+ ckpt_b = gr.Textbox(
+ label=i18n("Path to Model B"), value="", interactive=True
+ )
+ alpha_a = gr.Slider(
+ minimum=0,
+ maximum=1,
+ label=i18n("Weight (w) for Model A"),
+ value=0.5,
+ interactive=True,
+ )
+ with gr.Column():
+ sr_ = gr.Radio(
+ label=i18n("Target sample rate"),
+ choices=["40k", "48k"],
+ value="40k",
+ interactive=True,
+ )
+ if_f0_ = gr.Radio(
+ label=i18n("Whether the model has pitch guidance"),
+ choices=[i18n("Yes"), i18n("No")],
+ value=i18n("Yes"),
+ interactive=True,
+ )
+ info__ = gr.Textbox(
+ label=i18n("Model information to be placed"),
+ value="",
+ max_lines=8,
+ interactive=True,
+ )
+ with gr.Column():
+ name_to_save0 = gr.Textbox(
+ label=i18n("Saved model name (without extension)"),
+ value="",
+ max_lines=1,
+ interactive=True,
+ )
+ version_2 = gr.Radio(
+ label=i18n("Model architecture version"),
+ choices=["v1", "v2"],
+ value="v1",
+ interactive=True,
+ )
+ but6 = gr.Button(i18n("Fusion"), variant="primary")
+ with gr.Row():
+ info4 = gr.Textbox(label=i18n("Output information"), value="")
+ but6.click(
+ merge,
+ [
+ ckpt_a,
+ ckpt_b,
+ alpha_a,
+ sr_,
+ if_f0_,
+ info__,
+ name_to_save0,
+ version_2,
+ ],
+ info4,
+ api_name="ckpt_merge",
+ ) # def merge(path1,path2,alpha1,sr,f0,info):
+
+ gr.Markdown(
+ value=i18n(
+ "### Modify model information\n> Only supported for small model files extracted from the 'weights' folder."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ ckpt_path0 = gr.Textbox(
+ label=i18n("Path to Model"), value="", interactive=True
+ )
+ info_ = gr.Textbox(
+ label=i18n("Model information to be modified"),
+ value="",
+ max_lines=8,
+ interactive=True,
+ )
+ name_to_save1 = gr.Textbox(
+ label=i18n("Save file name (default: same as the source file)"),
+ value="",
+ max_lines=1,
+ interactive=True,
+ )
+ with gr.Column():
+ but7 = gr.Button(i18n("Modify"), variant="primary")
+ info5 = gr.Textbox(label=i18n("Output information"), value="")
+ but7.click(
+ change_info,
+ [ckpt_path0, info_, name_to_save1],
+ info5,
+ api_name="ckpt_modify",
+ )
+
+ gr.Markdown(
+ value=i18n(
+ "### View model information\n> Only supported for small model files extracted from the 'weights' folder."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ ckpt_path1 = gr.File(label=i18n("Path to Model"))
+ but8 = gr.Button(i18n("View"), variant="primary")
+ with gr.Column():
+ info6 = gr.Textbox(label=i18n("Output information"), value="")
+ but8.click(show_info, [ckpt_path1], info6, api_name="ckpt_show")
+
+ gr.Markdown(
+ value=i18n(
+ "### Model extraction\n> Enter the path of the large file model under the 'logs' folder.\n\nThis is useful if you want to stop training halfway and manually extract and save a small model file, or if you want to test an intermediate model."
+ )
+ )
+ with gr.Row():
+ with gr.Column():
+ ckpt_path2 = gr.Textbox(
+ label=i18n("Path to Model"),
+ value="E:\\codes\\py39\\logs\\mi-test_f0_48k\\G_23333.pth",
+ interactive=True,
+ )
+ save_name = gr.Textbox(
+ label=i18n("Save name"), value="", interactive=True
+ )
+ with gr.Row():
+ sr__ = gr.Radio(
+ label=i18n("Target sample rate"),
+ choices=["32k", "40k", "48k"],
+ value="40k",
+ interactive=True,
+ )
+ if_f0__ = gr.Radio(
+ label=i18n(
+ "Whether the model has pitch guidance (1: yes, 0: no)"
+ ),
+ choices=["1", "0"],
+ value="1",
+ interactive=True,
+ )
+ version_1 = gr.Radio(
+ label=i18n("Model architecture version"),
+ choices=["v1", "v2"],
+ value="v2",
+ interactive=True,
+ )
+ info___ = gr.Textbox(
+ label=i18n("Model information to be placed"),
+ value="",
+ max_lines=8,
+ interactive=True,
+ )
+ extauthor = gr.Textbox(
+ label=i18n("Model Author"),
+ value="",
+ max_lines=1,
+ interactive=True,
+ )
+ with gr.Column():
+ but9 = gr.Button(i18n("Extract"), variant="primary")
+ info7 = gr.Textbox(label=i18n("Output information"), value="")
+ ckpt_path2.change(
+ change_info_, [ckpt_path2], [sr__, if_f0__, version_1]
+ )
+ but9.click(
+ extract_small_model,
+ [
+ ckpt_path2,
+ save_name,
+ extauthor,
+ sr__,
+ if_f0__,
+ info___,
+ version_1,
+ ],
+ info7,
+ api_name="ckpt_extract",
+ )
+
+ with gr.TabItem(i18n("Export Onnx")):
+ with gr.Row():
+ ckpt_dir = gr.Textbox(
+ label=i18n("RVC Model Path"), value="", interactive=True
+ )
+ with gr.Row():
+ onnx_dir = gr.Textbox(
+ label=i18n("Onnx Export Path"), value="", interactive=True
+ )
+ with gr.Row():
+ infoOnnx = gr.Label(label="info")
+ with gr.Row():
+ butOnnx = gr.Button(i18n("Export Onnx Model"), variant="primary")
+ butOnnx.click(
+ export_onnx, [ckpt_dir, onnx_dir], infoOnnx, api_name="export_onnx"
+ )
+
+ tab_faq = i18n("FAQ (Frequently Asked Questions)")
+ with gr.TabItem(tab_faq):
+ try:
+ if tab_faq == "FAQ (Frequently Asked Questions)":
+ with open("docs/cn/faq.md", "r", encoding="utf8") as f:
+ info = f.read()
+ else:
+ with open("docs/en/faq_en.md", "r", encoding="utf8") as f:
+ info = f.read()
+ gr.Markdown(value=info)
+ except:
+ gr.Markdown(traceback.format_exc())
+
+try:
+ import signal
+
+ def cleanup(signum, frame):
+ signame = signal.Signals(signum).name
+ print(f"Got signal {signame} ({signum})")
+ app.close()
+ sys.exit(0)
+
+ signal.signal(signal.SIGINT, cleanup)
+ signal.signal(signal.SIGTERM, cleanup)
+ if config.global_link:
+ app.queue(max_size=1022).launch(share=True, max_threads=511)
+ else:
+ app.queue(max_size=1022).launch(
+ max_threads=511,
+ server_name="0.0.0.0",
+ inbrowser=not config.noautoopen,
+ server_port=config.listen_port,
+ quiet=True,
+ )
+except Exception as e:
+ logger.error(str(e))