diff --git a/.gitattributes b/.gitattributes
index a6344aac8c09253b3b630fb776ae94478aa0275b..1aecc638e2db2994100893e0155fb17d8c17d56e 100644
--- a/.gitattributes
+++ b/.gitattributes
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+GPT_SoVITS/text/ja_userdic/userdict.csv filter=lfs diff=lfs merge=lfs -text
diff --git a/GPT_SoVITS/TTS_infer_pack/TTS.py b/GPT_SoVITS/TTS_infer_pack/TTS.py
new file mode 100644
index 0000000000000000000000000000000000000000..1b7ad11a2f052c3cd0fed65802a448f18fbee2d4
--- /dev/null
+++ b/GPT_SoVITS/TTS_infer_pack/TTS.py
@@ -0,0 +1,1463 @@
+import gc
+import math
+import os
+import random
+import sys
+import time
+import traceback
+from copy import deepcopy
+
+import torchaudio
+from tqdm import tqdm
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+import os
+from typing import List, Tuple, Union
+
+import ffmpeg
+import librosa
+import numpy as np
+import torch
+import torch.nn.functional as F
+import yaml
+from AR.models.t2s_lightning_module import Text2SemanticLightningModule
+from BigVGAN.bigvgan import BigVGAN
+from feature_extractor.cnhubert import CNHubert
+from module.mel_processing import mel_spectrogram_torch, spectrogram_torch
+from module.models import SynthesizerTrn, SynthesizerTrnV3
+from peft import LoraConfig, get_peft_model
+from process_ckpt import get_sovits_version_from_path_fast, load_sovits_new
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+
+from tools.audio_sr import AP_BWE
+from tools.i18n.i18n import I18nAuto, scan_language_list
+from tools.my_utils import load_audio
+from TTS_infer_pack.text_segmentation_method import splits
+from TTS_infer_pack.TextPreprocessor import TextPreprocessor
+
+language = os.environ.get("language", "Auto")
+language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else language
+i18n = I18nAuto(language=language)
+
+
+spec_min = -12
+spec_max = 2
+
+
+def norm_spec(x):
+    return (x - spec_min) / (spec_max - spec_min) * 2 - 1
+
+
+def denorm_spec(x):
+    return (x + 1) / 2 * (spec_max - spec_min) + spec_min
+
+
+mel_fn = lambda x: mel_spectrogram_torch(
+    x,
+    **{
+        "n_fft": 1024,
+        "win_size": 1024,
+        "hop_size": 256,
+        "num_mels": 100,
+        "sampling_rate": 24000,
+        "fmin": 0,
+        "fmax": None,
+        "center": False,
+    },
+)
+
+
+def speed_change(input_audio: np.ndarray, speed: float, sr: int):
+    # 将 NumPy 数组转换为原始 PCM 流
+    raw_audio = input_audio.astype(np.int16).tobytes()
+
+    # 设置 ffmpeg 输入流
+    input_stream = ffmpeg.input("pipe:", format="s16le", acodec="pcm_s16le", ar=str(sr), ac=1)
+
+    # 变速处理
+    output_stream = input_stream.filter("atempo", speed)
+
+    # 输出流到管道
+    out, _ = output_stream.output("pipe:", format="s16le", acodec="pcm_s16le").run(
+        input=raw_audio, capture_stdout=True, capture_stderr=True
+    )
+
+    # 将管道输出解码为 NumPy 数组
+    processed_audio = np.frombuffer(out, np.int16)
+
+    return processed_audio
+
+
+resample_transform_dict = {}
+
+
+def resample(audio_tensor, sr0, device):
+    global resample_transform_dict
+    if sr0 not in resample_transform_dict:
+        resample_transform_dict[sr0] = torchaudio.transforms.Resample(sr0, 24000).to(device)
+    return resample_transform_dict[sr0](audio_tensor)
+
+
+class DictToAttrRecursive(dict):
+    def __init__(self, input_dict):
+        super().__init__(input_dict)
+        for key, value in input_dict.items():
+            if isinstance(value, dict):
+                value = DictToAttrRecursive(value)
+            self[key] = value
+            setattr(self, key, value)
+
+    def __getattr__(self, item):
+        try:
+            return self[item]
+        except KeyError:
+            raise AttributeError(f"Attribute {item} not found")
+
+    def __setattr__(self, key, value):
+        if isinstance(value, dict):
+            value = DictToAttrRecursive(value)
+        super(DictToAttrRecursive, self).__setitem__(key, value)
+        super().__setattr__(key, value)
+
+    def __delattr__(self, item):
+        try:
+            del self[item]
+        except KeyError:
+            raise AttributeError(f"Attribute {item} not found")
+
+
+class NO_PROMPT_ERROR(Exception):
+    pass
+
+
+# configs/tts_infer.yaml
+"""
+custom:
+  bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
+  cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
+  device: cpu
+  is_half: false
+  t2s_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt
+  vits_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth
+  version: v2
+default:
+  bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
+  cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
+  device: cpu
+  is_half: false
+  t2s_weights_path: GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt
+  vits_weights_path: GPT_SoVITS/pretrained_models/s2G488k.pth
+  version: v1
+default_v2:
+  bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
+  cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
+  device: cpu
+  is_half: false
+  t2s_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt
+  vits_weights_path: GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth
+  version: v2
+default_v3:
+  bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
+  cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base
+  device: cpu
+  is_half: false
+  t2s_weights_path: GPT_SoVITS/pretrained_models/s1v3.ckpt
+  vits_weights_path: GPT_SoVITS/pretrained_models/s2Gv3.pth
+  version: v3
+"""
+
+
+def set_seed(seed: int):
+    seed = int(seed)
+    seed = seed if seed != -1 else random.randint(0, 2**32 - 1)
+    print(f"Set seed to {seed}")
+    os.environ["PYTHONHASHSEED"] = str(seed)
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    try:
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed(seed)
+            torch.cuda.manual_seed_all(seed)
+            # torch.backends.cudnn.deterministic = True
+            # torch.backends.cudnn.benchmark = False
+            # torch.backends.cudnn.enabled = True
+            # 开启后会影响精度
+            torch.backends.cuda.matmul.allow_tf32 = False
+            torch.backends.cudnn.allow_tf32 = False
+    except:
+        pass
+    return seed
+
+
+class TTS_Config:
+    default_configs = {
+        "v1": {
+            "device": "cpu",
+            "is_half": False,
+            "version": "v1",
+            "t2s_weights_path": "GPT_SoVITS/pretrained_models/s1bert25hz-2kh-longer-epoch=68e-step=50232.ckpt",
+            "vits_weights_path": "GPT_SoVITS/pretrained_models/s2G488k.pth",
+            "cnhuhbert_base_path": "GPT_SoVITS/pretrained_models/chinese-hubert-base",
+            "bert_base_path": "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",
+        },
+        "v2": {
+            "device": "cpu",
+            "is_half": False,
+            "version": "v2",
+            "t2s_weights_path": "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt",
+            "vits_weights_path": "GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth",
+            "cnhuhbert_base_path": "GPT_SoVITS/pretrained_models/chinese-hubert-base",
+            "bert_base_path": "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",
+        },
+        "v3": {
+            "device": "cpu",
+            "is_half": False,
+            "version": "v3",
+            "t2s_weights_path": "GPT_SoVITS/pretrained_models/s1v3.ckpt",
+            "vits_weights_path": "GPT_SoVITS/pretrained_models/s2Gv3.pth",
+            "cnhuhbert_base_path": "GPT_SoVITS/pretrained_models/chinese-hubert-base",
+            "bert_base_path": "GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large",
+        },
+    }
+    configs: dict = None
+    v1_languages: list = ["auto", "en", "zh", "ja", "all_zh", "all_ja"]
+    v2_languages: list = ["auto", "auto_yue", "en", "zh", "ja", "yue", "ko", "all_zh", "all_ja", "all_yue", "all_ko"]
+    languages: list = v2_languages
+    # "all_zh",#全部按中文识别
+    # "en",#全部按英文识别#######不变
+    # "all_ja",#全部按日文识别
+    # "all_yue",#全部按中文识别
+    # "all_ko",#全部按韩文识别
+    # "zh",#按中英混合识别####不变
+    # "ja",#按日英混合识别####不变
+    # "yue",#按粤英混合识别####不变
+    # "ko",#按韩英混合识别####不变
+    # "auto",#多语种启动切分识别语种
+    # "auto_yue",#多语种启动切分识别语种
+
+    def __init__(self, configs: Union[dict, str] = None):
+        # 设置默认配置文件路径
+        configs_base_path: str = "GPT_SoVITS/configs/"
+        os.makedirs(configs_base_path, exist_ok=True)
+        self.configs_path: str = os.path.join(configs_base_path, "tts_infer.yaml")
+
+        if configs in ["", None]:
+            if not os.path.exists(self.configs_path):
+                self.save_configs()
+                print(f"Create default config file at {self.configs_path}")
+            configs: dict = deepcopy(self.default_configs)
+
+        if isinstance(configs, str):
+            self.configs_path = configs
+            configs: dict = self._load_configs(self.configs_path)
+
+        assert isinstance(configs, dict)
+        version = configs.get("version", "v2").lower()
+        assert version in ["v1", "v2", "v3"]
+        self.default_configs[version] = configs.get(version, self.default_configs[version])
+        self.configs: dict = configs.get("custom", deepcopy(self.default_configs[version]))
+
+        self.device = self.configs.get("device", torch.device("cpu"))
+        if "cuda" in str(self.device) and not torch.cuda.is_available():
+            print("Warning: CUDA is not available, set device to CPU.")
+            self.device = torch.device("cpu")
+
+        self.is_half = self.configs.get("is_half", False)
+        # if str(self.device) == "cpu" and self.is_half:
+        #     print(f"Warning: Half precision is not supported on CPU, set is_half to False.")
+        #     self.is_half = False
+
+        self.version = version
+        self.t2s_weights_path = self.configs.get("t2s_weights_path", None)
+        self.vits_weights_path = self.configs.get("vits_weights_path", None)
+        self.bert_base_path = self.configs.get("bert_base_path", None)
+        self.cnhuhbert_base_path = self.configs.get("cnhuhbert_base_path", None)
+        self.languages = self.v1_languages if self.version == "v1" else self.v2_languages
+
+        self.is_v3_synthesizer: bool = False
+
+        if (self.t2s_weights_path in [None, ""]) or (not os.path.exists(self.t2s_weights_path)):
+            self.t2s_weights_path = self.default_configs[version]["t2s_weights_path"]
+            print(f"fall back to default t2s_weights_path: {self.t2s_weights_path}")
+        if (self.vits_weights_path in [None, ""]) or (not os.path.exists(self.vits_weights_path)):
+            self.vits_weights_path = self.default_configs[version]["vits_weights_path"]
+            print(f"fall back to default vits_weights_path: {self.vits_weights_path}")
+        if (self.bert_base_path in [None, ""]) or (not os.path.exists(self.bert_base_path)):
+            self.bert_base_path = self.default_configs[version]["bert_base_path"]
+            print(f"fall back to default bert_base_path: {self.bert_base_path}")
+        if (self.cnhuhbert_base_path in [None, ""]) or (not os.path.exists(self.cnhuhbert_base_path)):
+            self.cnhuhbert_base_path = self.default_configs[version]["cnhuhbert_base_path"]
+            print(f"fall back to default cnhuhbert_base_path: {self.cnhuhbert_base_path}")
+        self.update_configs()
+
+        self.max_sec = None
+        self.hz: int = 50
+        self.semantic_frame_rate: str = "25hz"
+        self.segment_size: int = 20480
+        self.filter_length: int = 2048
+        self.sampling_rate: int = 32000
+        self.hop_length: int = 640
+        self.win_length: int = 2048
+        self.n_speakers: int = 300
+
+    def _load_configs(self, configs_path: str) -> dict:
+        if os.path.exists(configs_path):
+            ...
+        else:
+            print(i18n("路径不存在,使用默认配置"))
+            self.save_configs(configs_path)
+        with open(configs_path, "r", encoding="utf-8") as f:
+            configs = yaml.load(f, Loader=yaml.FullLoader)
+
+        return configs
+
+    def save_configs(self, configs_path: str = None) -> None:
+        configs = deepcopy(self.default_configs)
+        if self.configs is not None:
+            configs["custom"] = self.update_configs()
+
+        if configs_path is None:
+            configs_path = self.configs_path
+        with open(configs_path, "w") as f:
+            yaml.dump(configs, f)
+
+    def update_configs(self):
+        self.config = {
+            "device": str(self.device),
+            "is_half": self.is_half,
+            "version": self.version,
+            "t2s_weights_path": self.t2s_weights_path,
+            "vits_weights_path": self.vits_weights_path,
+            "bert_base_path": self.bert_base_path,
+            "cnhuhbert_base_path": self.cnhuhbert_base_path,
+        }
+        return self.config
+
+    def update_version(self, version: str) -> None:
+        self.version = version
+        self.languages = self.v1_languages if self.version == "v1" else self.v2_languages
+
+    def __str__(self):
+        self.configs = self.update_configs()
+        string = "TTS Config".center(100, "-") + "\n"
+        for k, v in self.configs.items():
+            string += f"{str(k).ljust(20)}: {str(v)}\n"
+        string += "-" * 100 + "\n"
+        return string
+
+    def __repr__(self):
+        return self.__str__()
+
+    def __hash__(self):
+        return hash(self.configs_path)
+
+    def __eq__(self, other):
+        return isinstance(other, TTS_Config) and self.configs_path == other.configs_path
+
+
+class TTS:
+    def __init__(self, configs: Union[dict, str, TTS_Config]):
+        if isinstance(configs, TTS_Config):
+            self.configs = configs
+        else:
+            self.configs: TTS_Config = TTS_Config(configs)
+
+        self.t2s_model: Text2SemanticLightningModule = None
+        self.vits_model: Union[SynthesizerTrn, SynthesizerTrnV3] = None
+        self.bert_tokenizer: AutoTokenizer = None
+        self.bert_model: AutoModelForMaskedLM = None
+        self.cnhuhbert_model: CNHubert = None
+        self.bigvgan_model: BigVGAN = None
+        self.sr_model: AP_BWE = None
+        self.sr_model_not_exist: bool = False
+
+        self._init_models()
+
+        self.text_preprocessor: TextPreprocessor = TextPreprocessor(
+            self.bert_model, self.bert_tokenizer, self.configs.device
+        )
+
+        self.prompt_cache: dict = {
+            "ref_audio_path": None,
+            "prompt_semantic": None,
+            "refer_spec": [],
+            "prompt_text": None,
+            "prompt_lang": None,
+            "phones": None,
+            "bert_features": None,
+            "norm_text": None,
+            "aux_ref_audio_paths": [],
+        }
+
+        self.stop_flag: bool = False
+        self.precision: torch.dtype = torch.float16 if self.configs.is_half else torch.float32
+
+    def _init_models(
+        self,
+    ):
+        self.init_t2s_weights(self.configs.t2s_weights_path)
+        self.init_vits_weights(self.configs.vits_weights_path)
+        self.init_bert_weights(self.configs.bert_base_path)
+        self.init_cnhuhbert_weights(self.configs.cnhuhbert_base_path)
+        # self.enable_half_precision(self.configs.is_half)
+
+    def init_cnhuhbert_weights(self, base_path: str):
+        print(f"Loading CNHuBERT weights from {base_path}")
+        self.cnhuhbert_model = CNHubert(base_path)
+        self.cnhuhbert_model = self.cnhuhbert_model.eval()
+        self.cnhuhbert_model = self.cnhuhbert_model.to(self.configs.device)
+        if self.configs.is_half and str(self.configs.device) != "cpu":
+            self.cnhuhbert_model = self.cnhuhbert_model.half()
+
+    def init_bert_weights(self, base_path: str):
+        print(f"Loading BERT weights from {base_path}")
+        self.bert_tokenizer = AutoTokenizer.from_pretrained(base_path)
+        self.bert_model = AutoModelForMaskedLM.from_pretrained(base_path)
+        self.bert_model = self.bert_model.eval()
+        self.bert_model = self.bert_model.to(self.configs.device)
+        if self.configs.is_half and str(self.configs.device) != "cpu":
+            self.bert_model = self.bert_model.half()
+
+    def init_vits_weights(self, weights_path: str):
+        self.configs.vits_weights_path = weights_path
+        version, model_version, if_lora_v3 = get_sovits_version_from_path_fast(weights_path)
+        path_sovits_v3 = self.configs.default_configs["v3"]["vits_weights_path"]
+
+        if if_lora_v3 == True and os.path.exists(path_sovits_v3) == False:
+            info = path_sovits_v3 + i18n("SoVITS V3 底模缺失，无法加载相应 LoRA 权重")
+            raise FileExistsError(info)
+
+        # dict_s2 = torch.load(weights_path, map_location=self.configs.device,weights_only=False)
+        dict_s2 = load_sovits_new(weights_path)
+        hps = dict_s2["config"]
+
+        hps["model"]["semantic_frame_rate"] = "25hz"
+        if "enc_p.text_embedding.weight" not in dict_s2["weight"]:
+            hps["model"]["version"] = "v2"  # v3model,v2sybomls
+        elif dict_s2["weight"]["enc_p.text_embedding.weight"].shape[0] == 322:
+            hps["model"]["version"] = "v1"
+        else:
+            hps["model"]["version"] = "v2"
+        # version = hps["model"]["version"]
+
+        self.configs.filter_length = hps["data"]["filter_length"]
+        self.configs.segment_size = hps["train"]["segment_size"]
+        self.configs.sampling_rate = hps["data"]["sampling_rate"]
+        self.configs.hop_length = hps["data"]["hop_length"]
+        self.configs.win_length = hps["data"]["win_length"]
+        self.configs.n_speakers = hps["data"]["n_speakers"]
+        self.configs.semantic_frame_rate = hps["model"]["semantic_frame_rate"]
+        kwargs = hps["model"]
+        # print(f"self.configs.sampling_rate:{self.configs.sampling_rate}")
+
+        self.configs.update_version(model_version)
+
+        # print(f"model_version:{model_version}")
+        # print(f'hps["model"]["version"]:{hps["model"]["version"]}')
+        if model_version != "v3":
+            vits_model = SynthesizerTrn(
+                self.configs.filter_length // 2 + 1,
+                self.configs.segment_size // self.configs.hop_length,
+                n_speakers=self.configs.n_speakers,
+                **kwargs,
+            )
+            self.configs.is_v3_synthesizer = False
+        else:
+            vits_model = SynthesizerTrnV3(
+                self.configs.filter_length // 2 + 1,
+                self.configs.segment_size // self.configs.hop_length,
+                n_speakers=self.configs.n_speakers,
+                **kwargs,
+            )
+            self.configs.is_v3_synthesizer = True
+            self.init_bigvgan()
+            if "pretrained" not in weights_path and hasattr(vits_model, "enc_q"):
+                del vits_model.enc_q
+
+        if if_lora_v3 == False:
+            print(
+                f"Loading VITS weights from {weights_path}. {vits_model.load_state_dict(dict_s2['weight'], strict=False)}"
+            )
+        else:
+            print(
+                f"Loading VITS pretrained weights from {weights_path}. {vits_model.load_state_dict(load_sovits_new(path_sovits_v3)['weight'], strict=False)}"
+            )
+            lora_rank = dict_s2["lora_rank"]
+            lora_config = LoraConfig(
+                target_modules=["to_k", "to_q", "to_v", "to_out.0"],
+                r=lora_rank,
+                lora_alpha=lora_rank,
+                init_lora_weights=True,
+            )
+            vits_model.cfm = get_peft_model(vits_model.cfm, lora_config)
+            print(
+                f"Loading LoRA weights from {weights_path}. {vits_model.load_state_dict(dict_s2['weight'], strict=False)}"
+            )
+
+            vits_model.cfm = vits_model.cfm.merge_and_unload()
+
+        vits_model = vits_model.to(self.configs.device)
+        vits_model = vits_model.eval()
+
+        self.vits_model = vits_model
+        if self.configs.is_half and str(self.configs.device) != "cpu":
+            self.vits_model = self.vits_model.half()
+
+    def init_t2s_weights(self, weights_path: str):
+        print(f"Loading Text2Semantic weights from {weights_path}")
+        self.configs.t2s_weights_path = weights_path
+        self.configs.save_configs()
+        self.configs.hz = 50
+        dict_s1 = torch.load(weights_path, map_location=self.configs.device)
+        config = dict_s1["config"]
+        self.configs.max_sec = config["data"]["max_sec"]
+        t2s_model = Text2SemanticLightningModule(config, "****", is_train=False)
+        t2s_model.load_state_dict(dict_s1["weight"])
+        t2s_model = t2s_model.to(self.configs.device)
+        t2s_model = t2s_model.eval()
+        self.t2s_model = t2s_model
+        if self.configs.is_half and str(self.configs.device) != "cpu":
+            self.t2s_model = self.t2s_model.half()
+
+    def init_bigvgan(self):
+        if self.bigvgan_model is not None:
+            return
+        self.bigvgan_model = BigVGAN.from_pretrained(
+            "%s/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x" % (now_dir,),
+            use_cuda_kernel=False,
+        )  # if True, RuntimeError: Ninja is required to load C++ extensions
+        # remove weight norm in the model and set to eval mode
+        self.bigvgan_model.remove_weight_norm()
+        self.bigvgan_model = self.bigvgan_model.eval()
+        if self.configs.is_half == True:
+            self.bigvgan_model = self.bigvgan_model.half().to(self.configs.device)
+        else:
+            self.bigvgan_model = self.bigvgan_model.to(self.configs.device)
+
+    def init_sr_model(self):
+        if self.sr_model is not None:
+            return
+        try:
+            self.sr_model: AP_BWE = AP_BWE(self.configs.device, DictToAttrRecursive)
+            self.sr_model_not_exist = False
+        except FileNotFoundError:
+            print(i18n("你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好"))
+            self.sr_model_not_exist = True
+
+    def enable_half_precision(self, enable: bool = True, save: bool = True):
+        """
+        To enable half precision for the TTS model.
+        Args:
+            enable: bool, whether to enable half precision.
+
+        """
+        if str(self.configs.device) == "cpu" and enable:
+            print("Half precision is not supported on CPU.")
+            return
+
+        self.configs.is_half = enable
+        self.precision = torch.float16 if enable else torch.float32
+        if save:
+            self.configs.save_configs()
+        if enable:
+            if self.t2s_model is not None:
+                self.t2s_model = self.t2s_model.half()
+            if self.vits_model is not None:
+                self.vits_model = self.vits_model.half()
+            if self.bert_model is not None:
+                self.bert_model = self.bert_model.half()
+            if self.cnhuhbert_model is not None:
+                self.cnhuhbert_model = self.cnhuhbert_model.half()
+            if self.bigvgan_model is not None:
+                self.bigvgan_model = self.bigvgan_model.half()
+        else:
+            if self.t2s_model is not None:
+                self.t2s_model = self.t2s_model.float()
+            if self.vits_model is not None:
+                self.vits_model = self.vits_model.float()
+            if self.bert_model is not None:
+                self.bert_model = self.bert_model.float()
+            if self.cnhuhbert_model is not None:
+                self.cnhuhbert_model = self.cnhuhbert_model.float()
+            if self.bigvgan_model is not None:
+                self.bigvgan_model = self.bigvgan_model.float()
+
+    def set_device(self, device: torch.device, save: bool = True):
+        """
+        To set the device for all models.
+        Args:
+            device: torch.device, the device to use for all models.
+        """
+        self.configs.device = device
+        if save:
+            self.configs.save_configs()
+        if self.t2s_model is not None:
+            self.t2s_model = self.t2s_model.to(device)
+        if self.vits_model is not None:
+            self.vits_model = self.vits_model.to(device)
+        if self.bert_model is not None:
+            self.bert_model = self.bert_model.to(device)
+        if self.cnhuhbert_model is not None:
+            self.cnhuhbert_model = self.cnhuhbert_model.to(device)
+        if self.bigvgan_model is not None:
+            self.bigvgan_model = self.bigvgan_model.to(device)
+        if self.sr_model is not None:
+            self.sr_model = self.sr_model.to(device)
+
+    def set_ref_audio(self, ref_audio_path: str):
+        """
+        To set the reference audio for the TTS model,
+            including the prompt_semantic and refer_spepc.
+        Args:
+            ref_audio_path: str, the path of the reference audio.
+        """
+        self._set_prompt_semantic(ref_audio_path)
+        self._set_ref_spec(ref_audio_path)
+        self._set_ref_audio_path(ref_audio_path)
+
+    def _set_ref_audio_path(self, ref_audio_path):
+        self.prompt_cache["ref_audio_path"] = ref_audio_path
+
+    def _set_ref_spec(self, ref_audio_path):
+        spec = self._get_ref_spec(ref_audio_path)
+        if self.prompt_cache["refer_spec"] in [[], None]:
+            self.prompt_cache["refer_spec"] = [spec]
+        else:
+            self.prompt_cache["refer_spec"][0] = spec
+
+    def _get_ref_spec(self, ref_audio_path):
+        raw_audio, raw_sr = torchaudio.load(ref_audio_path)
+        raw_audio = raw_audio.to(self.configs.device).float()
+        self.prompt_cache["raw_audio"] = raw_audio
+        self.prompt_cache["raw_sr"] = raw_sr
+
+        audio = load_audio(ref_audio_path, int(self.configs.sampling_rate))
+        audio = torch.FloatTensor(audio)
+        maxx = audio.abs().max()
+        if maxx > 1:
+            audio /= min(2, maxx)
+        audio_norm = audio
+        audio_norm = audio_norm.unsqueeze(0)
+        spec = spectrogram_torch(
+            audio_norm,
+            self.configs.filter_length,
+            self.configs.sampling_rate,
+            self.configs.hop_length,
+            self.configs.win_length,
+            center=False,
+        )
+        spec = spec.to(self.configs.device)
+        if self.configs.is_half:
+            spec = spec.half()
+        return spec
+
+    def _set_prompt_semantic(self, ref_wav_path: str):
+        zero_wav = np.zeros(
+            int(self.configs.sampling_rate * 0.3),
+            dtype=np.float16 if self.configs.is_half else np.float32,
+        )
+        with torch.no_grad():
+            wav16k, sr = librosa.load(ref_wav_path, sr=16000)
+            if wav16k.shape[0] > 160000 or wav16k.shape[0] < 48000:
+                raise OSError(i18n("参考音频在3~10秒范围外，请更换！"))
+            wav16k = torch.from_numpy(wav16k)
+            zero_wav_torch = torch.from_numpy(zero_wav)
+            wav16k = wav16k.to(self.configs.device)
+            zero_wav_torch = zero_wav_torch.to(self.configs.device)
+            if self.configs.is_half:
+                wav16k = wav16k.half()
+                zero_wav_torch = zero_wav_torch.half()
+
+            wav16k = torch.cat([wav16k, zero_wav_torch])
+            hubert_feature = self.cnhuhbert_model.model(wav16k.unsqueeze(0))["last_hidden_state"].transpose(
+                1, 2
+            )  # .float()
+            codes = self.vits_model.extract_latent(hubert_feature)
+
+            prompt_semantic = codes[0, 0].to(self.configs.device)
+            self.prompt_cache["prompt_semantic"] = prompt_semantic
+
+    def batch_sequences(self, sequences: List[torch.Tensor], axis: int = 0, pad_value: int = 0, max_length: int = None):
+        seq = sequences[0]
+        ndim = seq.dim()
+        if axis < 0:
+            axis += ndim
+        dtype: torch.dtype = seq.dtype
+        pad_value = torch.tensor(pad_value, dtype=dtype)
+        seq_lengths = [seq.shape[axis] for seq in sequences]
+        if max_length is None:
+            max_length = max(seq_lengths)
+        else:
+            max_length = max(seq_lengths) if max_length < max(seq_lengths) else max_length
+
+        padded_sequences = []
+        for seq, length in zip(sequences, seq_lengths):
+            padding = [0] * axis + [0, max_length - length] + [0] * (ndim - axis - 1)
+            padded_seq = torch.nn.functional.pad(seq, padding, value=pad_value)
+            padded_sequences.append(padded_seq)
+        batch = torch.stack(padded_sequences)
+        return batch
+
+    def to_batch(
+        self,
+        data: list,
+        prompt_data: dict = None,
+        batch_size: int = 5,
+        threshold: float = 0.75,
+        split_bucket: bool = True,
+        device: torch.device = torch.device("cpu"),
+        precision: torch.dtype = torch.float32,
+    ):
+        _data: list = []
+        index_and_len_list = []
+        for idx, item in enumerate(data):
+            norm_text_len = len(item["norm_text"])
+            index_and_len_list.append([idx, norm_text_len])
+
+        batch_index_list = []
+        if split_bucket:
+            index_and_len_list.sort(key=lambda x: x[1])
+            index_and_len_list = np.array(index_and_len_list, dtype=np.int64)
+
+            batch_index_list_len = 0
+            pos = 0
+            while pos < index_and_len_list.shape[0]:
+                # batch_index_list.append(index_and_len_list[pos:min(pos+batch_size,len(index_and_len_list))])
+                pos_end = min(pos + batch_size, index_and_len_list.shape[0])
+                while pos < pos_end:
+                    batch = index_and_len_list[pos:pos_end, 1].astype(np.float32)
+                    score = batch[(pos_end - pos) // 2] / (batch.mean() + 1e-8)
+                    if (score >= threshold) or (pos_end - pos == 1):
+                        batch_index = index_and_len_list[pos:pos_end, 0].tolist()
+                        batch_index_list_len += len(batch_index)
+                        batch_index_list.append(batch_index)
+                        pos = pos_end
+                        break
+                    pos_end = pos_end - 1
+
+            assert batch_index_list_len == len(data)
+
+        else:
+            for i in range(len(data)):
+                if i % batch_size == 0:
+                    batch_index_list.append([])
+                batch_index_list[-1].append(i)
+
+        for batch_idx, index_list in enumerate(batch_index_list):
+            item_list = [data[idx] for idx in index_list]
+            phones_list = []
+            phones_len_list = []
+            # bert_features_list = []
+            all_phones_list = []
+            all_phones_len_list = []
+            all_bert_features_list = []
+            norm_text_batch = []
+            all_bert_max_len = 0
+            all_phones_max_len = 0
+            for item in item_list:
+                if prompt_data is not None:
+                    all_bert_features = torch.cat([prompt_data["bert_features"], item["bert_features"]], 1).to(
+                        dtype=precision, device=device
+                    )
+                    all_phones = torch.LongTensor(prompt_data["phones"] + item["phones"]).to(device)
+                    phones = torch.LongTensor(item["phones"]).to(device)
+                    # norm_text = prompt_data["norm_text"]+item["norm_text"]
+                else:
+                    all_bert_features = item["bert_features"].to(dtype=precision, device=device)
+                    phones = torch.LongTensor(item["phones"]).to(device)
+                    all_phones = phones
+                    # norm_text = item["norm_text"]
+
+                all_bert_max_len = max(all_bert_max_len, all_bert_features.shape[-1])
+                all_phones_max_len = max(all_phones_max_len, all_phones.shape[-1])
+
+                phones_list.append(phones)
+                phones_len_list.append(phones.shape[-1])
+                all_phones_list.append(all_phones)
+                all_phones_len_list.append(all_phones.shape[-1])
+                all_bert_features_list.append(all_bert_features)
+                norm_text_batch.append(item["norm_text"])
+
+            phones_batch = phones_list
+            all_phones_batch = all_phones_list
+            all_bert_features_batch = all_bert_features_list
+
+            max_len = max(all_bert_max_len, all_phones_max_len)
+            # phones_batch = self.batch_sequences(phones_list, axis=0, pad_value=0, max_length=max_len)
+            #### 直接对phones和bert_features进行pad。（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略）
+            # all_phones_batch = self.batch_sequences(all_phones_list, axis=0, pad_value=0, max_length=max_len)
+            # all_bert_features_batch = all_bert_features_list
+            # all_bert_features_batch = torch.zeros((len(all_bert_features_list), 1024, max_len), dtype=precision, device=device)
+            # for idx, item in enumerate(all_bert_features_list):
+            #     all_bert_features_batch[idx, :, : item.shape[-1]] = item
+
+            # #### 先对phones进行embedding、对bert_features进行project，再pad到相同长度，（padding策略会影响T2S模型生成的结果，但不直接影响复读概率。影响复读概率的主要因素是mask的策略）
+            # all_phones_list = [self.t2s_model.model.ar_text_embedding(item.to(self.t2s_model.device)) for item in all_phones_list]
+            # all_phones_list = [F.pad(item,(0,0,0,max_len-item.shape[0]),value=0) for item in all_phones_list]
+            # all_phones_batch = torch.stack(all_phones_list, dim=0)
+
+            # all_bert_features_list = [self.t2s_model.model.bert_proj(item.to(self.t2s_model.device).transpose(0, 1)) for item in all_bert_features_list]
+            # all_bert_features_list = [F.pad(item,(0,0,0,max_len-item.shape[0]), value=0) for item in all_bert_features_list]
+            # all_bert_features_batch = torch.stack(all_bert_features_list, dim=0)
+
+            batch = {
+                "phones": phones_batch,
+                "phones_len": torch.LongTensor(phones_len_list).to(device),
+                "all_phones": all_phones_batch,
+                "all_phones_len": torch.LongTensor(all_phones_len_list).to(device),
+                "all_bert_features": all_bert_features_batch,
+                "norm_text": norm_text_batch,
+                "max_len": max_len,
+            }
+            _data.append(batch)
+
+        return _data, batch_index_list
+
+    def recovery_order(self, data: list, batch_index_list: list) -> list:
+        """
+        Recovery the order of the audio according to the batch_index_list.
+
+        Args:
+            data (List[list(torch.Tensor)]): the out of order audio .
+            batch_index_list (List[list[int]]): the batch index list.
+
+        Returns:
+            list (List[torch.Tensor]): the data in the original order.
+        """
+        length = len(sum(batch_index_list, []))
+        _data = [None] * length
+        for i, index_list in enumerate(batch_index_list):
+            for j, index in enumerate(index_list):
+                _data[index] = data[i][j]
+        return _data
+
+    def stop(
+        self,
+    ):
+        """
+        Stop the inference process.
+        """
+        self.stop_flag = True
+
+    @torch.no_grad()
+    def run(self, inputs: dict):
+        """
+        Text to speech inference.
+
+        Args:
+            inputs (dict):
+                {
+                    "text": "",                   # str.(required) text to be synthesized
+                    "text_lang: "",               # str.(required) language of the text to be synthesized
+                    "ref_audio_path": "",         # str.(required) reference audio path
+                    "aux_ref_audio_paths": [],    # list.(optional) auxiliary reference audio paths for multi-speaker tone fusion
+                    "prompt_text": "",            # str.(optional) prompt text for the reference audio
+                    "prompt_lang": "",            # str.(required) language of the prompt text for the reference audio
+                    "top_k": 5,                   # int. top k sampling
+                    "top_p": 1,                   # float. top p sampling
+                    "temperature": 1,             # float. temperature for sampling
+                    "text_split_method": "cut0",  # str. text split method, see text_segmentation_method.py for details.
+                    "batch_size": 1,              # int. batch size for inference
+                    "batch_threshold": 0.75,      # float. threshold for batch splitting.
+                    "split_bucket: True,          # bool. whether to split the batch into multiple buckets.
+                    "return_fragment": False,     # bool. step by step return the audio fragment.
+                    "speed_factor":1.0,           # float. control the speed of the synthesized audio.
+                    "fragment_interval":0.3,      # float. to control the interval of the audio fragment.
+                    "seed": -1,                   # int. random seed for reproducibility.
+                    "parallel_infer": True,       # bool. whether to use parallel inference.
+                    "repetition_penalty": 1.35    # float. repetition penalty for T2S model.
+                    "sample_steps": 32,           # int. number of sampling steps for VITS model V3.
+                    "super_sampling": False,       # bool. whether to use super-sampling for audio when using VITS model V3.
+                }
+        returns:
+            Tuple[int, np.ndarray]: sampling rate and audio data.
+        """
+        ########## variables initialization ###########
+        self.stop_flag: bool = False
+        text: str = inputs.get("text", "")
+        text_lang: str = inputs.get("text_lang", "")
+        ref_audio_path: str = inputs.get("ref_audio_path", "")
+        aux_ref_audio_paths: list = inputs.get("aux_ref_audio_paths", [])
+        prompt_text: str = inputs.get("prompt_text", "")
+        prompt_lang: str = inputs.get("prompt_lang", "")
+        top_k: int = inputs.get("top_k", 5)
+        top_p: float = inputs.get("top_p", 1)
+        temperature: float = inputs.get("temperature", 1)
+        text_split_method: str = inputs.get("text_split_method", "cut0")
+        batch_size = inputs.get("batch_size", 1)
+        batch_threshold = inputs.get("batch_threshold", 0.75)
+        speed_factor = inputs.get("speed_factor", 1.0)
+        split_bucket = inputs.get("split_bucket", True)
+        return_fragment = inputs.get("return_fragment", False)
+        fragment_interval = inputs.get("fragment_interval", 0.3)
+        seed = inputs.get("seed", -1)
+        seed = -1 if seed in ["", None] else seed
+        actual_seed = set_seed(seed)
+        parallel_infer = inputs.get("parallel_infer", True)
+        repetition_penalty = inputs.get("repetition_penalty", 1.35)
+        sample_steps = inputs.get("sample_steps", 32)
+        super_sampling = inputs.get("super_sampling", False)
+
+        if parallel_infer:
+            print(i18n("并行推理模式已开启"))
+            self.t2s_model.model.infer_panel = self.t2s_model.model.infer_panel_batch_infer
+        else:
+            print(i18n("并行推理模式已关闭"))
+            self.t2s_model.model.infer_panel = self.t2s_model.model.infer_panel_naive_batched
+
+        if return_fragment:
+            print(i18n("分段返回模式已开启"))
+            if split_bucket:
+                split_bucket = False
+                print(i18n("分段返回模式不支持分桶处理，已自动关闭分桶处理"))
+
+        if split_bucket and speed_factor == 1.0 and not (self.configs.is_v3_synthesizer and parallel_infer):
+            print(i18n("分桶处理模式已开启"))
+        elif speed_factor != 1.0:
+            print(i18n("语速调节不支持分桶处理，已自动关闭分桶处理"))
+            split_bucket = False
+        elif self.configs.is_v3_synthesizer and parallel_infer:
+            print(i18n("当开启并行推理模式时，SoVits V3模型不支持分桶处理，已自动关闭分桶处理"))
+            split_bucket = False
+        else:
+            print(i18n("分桶处理模式已关闭"))
+
+        if fragment_interval < 0.01:
+            fragment_interval = 0.01
+            print(i18n("分段间隔过小，已自动设置为0.01"))
+
+        no_prompt_text = False
+        if prompt_text in [None, ""]:
+            no_prompt_text = True
+
+        assert text_lang in self.configs.languages
+        if not no_prompt_text:
+            assert prompt_lang in self.configs.languages
+
+        if no_prompt_text and self.configs.is_v3_synthesizer:
+            raise NO_PROMPT_ERROR("prompt_text cannot be empty when using SoVITS_V3")
+
+        if ref_audio_path in [None, ""] and (
+            (self.prompt_cache["prompt_semantic"] is None) or (self.prompt_cache["refer_spec"] in [None, []])
+        ):
+            raise ValueError(
+                "ref_audio_path cannot be empty, when the reference audio is not set using set_ref_audio()"
+            )
+
+        ###### setting reference audio and prompt text preprocessing ########
+        t0 = time.perf_counter()
+        if (ref_audio_path is not None) and (ref_audio_path != self.prompt_cache["ref_audio_path"]):
+            if not os.path.exists(ref_audio_path):
+                raise ValueError(f"{ref_audio_path} not exists")
+            self.set_ref_audio(ref_audio_path)
+
+        aux_ref_audio_paths = aux_ref_audio_paths if aux_ref_audio_paths is not None else []
+        paths = set(aux_ref_audio_paths) & set(self.prompt_cache["aux_ref_audio_paths"])
+        if not (len(list(paths)) == len(aux_ref_audio_paths) == len(self.prompt_cache["aux_ref_audio_paths"])):
+            self.prompt_cache["aux_ref_audio_paths"] = aux_ref_audio_paths
+            self.prompt_cache["refer_spec"] = [self.prompt_cache["refer_spec"][0]]
+            for path in aux_ref_audio_paths:
+                if path in [None, ""]:
+                    continue
+                if not os.path.exists(path):
+                    print(i18n("音频文件不存在，跳过："), path)
+                    continue
+                self.prompt_cache["refer_spec"].append(self._get_ref_spec(path))
+
+        if not no_prompt_text:
+            prompt_text = prompt_text.strip("\n")
+            if prompt_text[-1] not in splits:
+                prompt_text += "。" if prompt_lang != "en" else "."
+            print(i18n("实际输入的参考文本:"), prompt_text)
+            if self.prompt_cache["prompt_text"] != prompt_text:
+                phones, bert_features, norm_text = self.text_preprocessor.segment_and_extract_feature_for_text(
+                    prompt_text, prompt_lang, self.configs.version
+                )
+                self.prompt_cache["prompt_text"] = prompt_text
+                self.prompt_cache["prompt_lang"] = prompt_lang
+                self.prompt_cache["phones"] = phones
+                self.prompt_cache["bert_features"] = bert_features
+                self.prompt_cache["norm_text"] = norm_text
+
+        ###### text preprocessing ########
+        t1 = time.perf_counter()
+        data: list = None
+        if not return_fragment:
+            data = self.text_preprocessor.preprocess(text, text_lang, text_split_method, self.configs.version)
+            if len(data) == 0:
+                yield 16000, np.zeros(int(16000), dtype=np.int16)
+                return
+
+            batch_index_list: list = None
+            data, batch_index_list = self.to_batch(
+                data,
+                prompt_data=self.prompt_cache if not no_prompt_text else None,
+                batch_size=batch_size,
+                threshold=batch_threshold,
+                split_bucket=split_bucket,
+                device=self.configs.device,
+                precision=self.precision,
+            )
+        else:
+            print(f"############ {i18n('切分文本')} ############")
+            texts = self.text_preprocessor.pre_seg_text(text, text_lang, text_split_method)
+            data = []
+            for i in range(len(texts)):
+                if i % batch_size == 0:
+                    data.append([])
+                data[-1].append(texts[i])
+
+            def make_batch(batch_texts):
+                batch_data = []
+                print(f"############ {i18n('提取文本Bert特征')} ############")
+                for text in tqdm(batch_texts):
+                    phones, bert_features, norm_text = self.text_preprocessor.segment_and_extract_feature_for_text(
+                        text, text_lang, self.configs.version
+                    )
+                    if phones is None:
+                        continue
+                    res = {
+                        "phones": phones,
+                        "bert_features": bert_features,
+                        "norm_text": norm_text,
+                    }
+                    batch_data.append(res)
+                if len(batch_data) == 0:
+                    return None
+                batch, _ = self.to_batch(
+                    batch_data,
+                    prompt_data=self.prompt_cache if not no_prompt_text else None,
+                    batch_size=batch_size,
+                    threshold=batch_threshold,
+                    split_bucket=False,
+                    device=self.configs.device,
+                    precision=self.precision,
+                )
+                return batch[0]
+
+        t2 = time.perf_counter()
+        try:
+            print("############ 推理 ############")
+            ###### inference ######
+            t_34 = 0.0
+            t_45 = 0.0
+            audio = []
+            output_sr = self.configs.sampling_rate if not self.configs.is_v3_synthesizer else 24000
+            for item in data:
+                t3 = time.perf_counter()
+                if return_fragment:
+                    item = make_batch(item)
+                    if item is None:
+                        continue
+
+                batch_phones: List[torch.LongTensor] = item["phones"]
+                # batch_phones:torch.LongTensor = item["phones"]
+                batch_phones_len: torch.LongTensor = item["phones_len"]
+                all_phoneme_ids: torch.LongTensor = item["all_phones"]
+                all_phoneme_lens: torch.LongTensor = item["all_phones_len"]
+                all_bert_features: torch.LongTensor = item["all_bert_features"]
+                norm_text: str = item["norm_text"]
+                max_len = item["max_len"]
+
+                print(i18n("前端处理后的文本(每句):"), norm_text)
+                if no_prompt_text:
+                    prompt = None
+                else:
+                    prompt = (
+                        self.prompt_cache["prompt_semantic"].expand(len(all_phoneme_ids), -1).to(self.configs.device)
+                    )
+
+                print(f"############ {i18n('预测语义Token')} ############")
+                pred_semantic_list, idx_list = self.t2s_model.model.infer_panel(
+                    all_phoneme_ids,
+                    all_phoneme_lens,
+                    prompt,
+                    all_bert_features,
+                    # prompt_phone_len=ph_offset,
+                    top_k=top_k,
+                    top_p=top_p,
+                    temperature=temperature,
+                    early_stop_num=self.configs.hz * self.configs.max_sec,
+                    max_len=max_len,
+                    repetition_penalty=repetition_penalty,
+                )
+                t4 = time.perf_counter()
+                t_34 += t4 - t3
+
+                refer_audio_spec: torch.Tensor = [
+                    item.to(dtype=self.precision, device=self.configs.device)
+                    for item in self.prompt_cache["refer_spec"]
+                ]
+
+                batch_audio_fragment = []
+
+                # ## vits并行推理 method 1
+                # pred_semantic_list = [item[-idx:] for item, idx in zip(pred_semantic_list, idx_list)]
+                # pred_semantic_len = torch.LongTensor([item.shape[0] for item in pred_semantic_list]).to(self.configs.device)
+                # pred_semantic = self.batch_sequences(pred_semantic_list, axis=0, pad_value=0).unsqueeze(0)
+                # max_len = 0
+                # for i in range(0, len(batch_phones)):
+                #     max_len = max(max_len, batch_phones[i].shape[-1])
+                # batch_phones = self.batch_sequences(batch_phones, axis=0, pad_value=0, max_length=max_len)
+                # batch_phones = batch_phones.to(self.configs.device)
+                # batch_audio_fragment = (self.vits_model.batched_decode(
+                #         pred_semantic, pred_semantic_len, batch_phones, batch_phones_len,refer_audio_spec
+                #     ))
+                print(f"############ {i18n('合成音频')} ############")
+                if not self.configs.is_v3_synthesizer:
+                    if speed_factor == 1.0:
+                        print(f"{i18n('并行合成中')}...")
+                        # ## vits并行推理 method 2
+                        pred_semantic_list = [item[-idx:] for item, idx in zip(pred_semantic_list, idx_list)]
+                        upsample_rate = math.prod(self.vits_model.upsample_rates)
+                        audio_frag_idx = [
+                            pred_semantic_list[i].shape[0] * 2 * upsample_rate
+                            for i in range(0, len(pred_semantic_list))
+                        ]
+                        audio_frag_end_idx = [sum(audio_frag_idx[: i + 1]) for i in range(0, len(audio_frag_idx))]
+                        all_pred_semantic = (
+                            torch.cat(pred_semantic_list).unsqueeze(0).unsqueeze(0).to(self.configs.device)
+                        )
+                        _batch_phones = torch.cat(batch_phones).unsqueeze(0).to(self.configs.device)
+                        _batch_audio_fragment = self.vits_model.decode(
+                            all_pred_semantic, _batch_phones, refer_audio_spec, speed=speed_factor
+                        ).detach()[0, 0, :]
+                        audio_frag_end_idx.insert(0, 0)
+                        batch_audio_fragment = [
+                            _batch_audio_fragment[audio_frag_end_idx[i - 1] : audio_frag_end_idx[i]]
+                            for i in range(1, len(audio_frag_end_idx))
+                        ]
+                    else:
+                        # ## vits串行推理
+                        for i, idx in enumerate(tqdm(idx_list)):
+                            phones = batch_phones[i].unsqueeze(0).to(self.configs.device)
+                            _pred_semantic = (
+                                pred_semantic_list[i][-idx:].unsqueeze(0).unsqueeze(0)
+                            )  # .unsqueeze(0)#mq要多unsqueeze一次
+                            audio_fragment = self.vits_model.decode(
+                                _pred_semantic, phones, refer_audio_spec, speed=speed_factor
+                            ).detach()[0, 0, :]
+                            batch_audio_fragment.append(audio_fragment)  ###试试重建不带上prompt部分
+                else:
+                    if parallel_infer:
+                        print(f"{i18n('并行合成中')}...")
+                        audio_fragments = self.v3_synthesis_batched_infer(
+                            idx_list, pred_semantic_list, batch_phones, speed=speed_factor, sample_steps=sample_steps
+                        )
+                        batch_audio_fragment.extend(audio_fragments)
+                    else:
+                        for i, idx in enumerate(tqdm(idx_list)):
+                            phones = batch_phones[i].unsqueeze(0).to(self.configs.device)
+                            _pred_semantic = (
+                                pred_semantic_list[i][-idx:].unsqueeze(0).unsqueeze(0)
+                            )  # .unsqueeze(0)#mq要多unsqueeze一次
+                            audio_fragment = self.v3_synthesis(
+                                _pred_semantic, phones, speed=speed_factor, sample_steps=sample_steps
+                            )
+                            batch_audio_fragment.append(audio_fragment)
+
+                t5 = time.perf_counter()
+                t_45 += t5 - t4
+                if return_fragment:
+                    print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t4 - t3, t5 - t4))
+                    yield self.audio_postprocess(
+                        [batch_audio_fragment],
+                        output_sr,
+                        None,
+                        speed_factor,
+                        False,
+                        fragment_interval,
+                        super_sampling if self.configs.is_v3_synthesizer else False,
+                    )
+                else:
+                    audio.append(batch_audio_fragment)
+
+                if self.stop_flag:
+                    yield 16000, np.zeros(int(16000), dtype=np.int16)
+                    return
+
+            if not return_fragment:
+                print("%.3f\t%.3f\t%.3f\t%.3f" % (t1 - t0, t2 - t1, t_34, t_45))
+                if len(audio) == 0:
+                    yield 16000, np.zeros(int(16000), dtype=np.int16)
+                    return
+                yield self.audio_postprocess(
+                    audio,
+                    output_sr,
+                    batch_index_list,
+                    speed_factor,
+                    split_bucket,
+                    fragment_interval,
+                    super_sampling if self.configs.is_v3_synthesizer else False,
+                )
+
+        except Exception as e:
+            traceback.print_exc()
+            # 必须返回一个空音频, 否则会导致显存不释放。
+            yield 16000, np.zeros(int(16000), dtype=np.int16)
+            # 重置模型, 否则会导致显存释放不完全。
+            del self.t2s_model
+            del self.vits_model
+            self.t2s_model = None
+            self.vits_model = None
+            self.init_t2s_weights(self.configs.t2s_weights_path)
+            self.init_vits_weights(self.configs.vits_weights_path)
+            raise e
+        finally:
+            self.empty_cache()
+
+    def empty_cache(self):
+        try:
+            gc.collect()  # 触发gc的垃圾回收。避免内存一直增长。
+            if "cuda" in str(self.configs.device):
+                torch.cuda.empty_cache()
+            elif str(self.configs.device) == "mps":
+                torch.mps.empty_cache()
+        except:
+            pass
+
+    def audio_postprocess(
+        self,
+        audio: List[torch.Tensor],
+        sr: int,
+        batch_index_list: list = None,
+        speed_factor: float = 1.0,
+        split_bucket: bool = True,
+        fragment_interval: float = 0.3,
+        super_sampling: bool = False,
+    ) -> Tuple[int, np.ndarray]:
+        zero_wav = torch.zeros(
+            int(self.configs.sampling_rate * fragment_interval), dtype=self.precision, device=self.configs.device
+        )
+
+        for i, batch in enumerate(audio):
+            for j, audio_fragment in enumerate(batch):
+                max_audio = torch.abs(audio_fragment).max()  # 简单防止16bit爆音
+                if max_audio > 1:
+                    audio_fragment /= max_audio
+                audio_fragment: torch.Tensor = torch.cat([audio_fragment, zero_wav], dim=0)
+                audio[i][j] = audio_fragment
+
+        if split_bucket:
+            audio = self.recovery_order(audio, batch_index_list)
+        else:
+            # audio = [item for batch in audio for item in batch]
+            audio = sum(audio, [])
+
+        audio = torch.cat(audio, dim=0)
+
+        if super_sampling:
+            print(f"############ {i18n('音频超采样')} ############")
+            t1 = time.perf_counter()
+            self.init_sr_model()
+            if not self.sr_model_not_exist:
+                audio, sr = self.sr_model(audio.unsqueeze(0), sr)
+                max_audio = np.abs(audio).max()
+                if max_audio > 1:
+                    audio /= max_audio
+            t2 = time.perf_counter()
+            print(f"超采样用时：{t2 - t1:.3f}s")
+        else:
+            audio = audio.cpu().numpy()
+
+        audio = (audio * 32768).astype(np.int16)
+
+        # try:
+        #     if speed_factor != 1.0:
+        #         audio = speed_change(audio, speed=speed_factor, sr=int(sr))
+        # except Exception as e:
+        #     print(f"Failed to change speed of audio: \n{e}")
+
+        return sr, audio
+
+    def v3_synthesis(
+        self, semantic_tokens: torch.Tensor, phones: torch.Tensor, speed: float = 1.0, sample_steps: int = 32
+    ):
+        prompt_semantic_tokens = self.prompt_cache["prompt_semantic"].unsqueeze(0).unsqueeze(0).to(self.configs.device)
+        prompt_phones = torch.LongTensor(self.prompt_cache["phones"]).unsqueeze(0).to(self.configs.device)
+        refer_audio_spec = self.prompt_cache["refer_spec"][0].to(dtype=self.precision, device=self.configs.device)
+
+        fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
+        ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]
+        ref_sr = self.prompt_cache["raw_sr"]
+        ref_audio = ref_audio.to(self.configs.device).float()
+        if ref_audio.shape[0] == 2:
+            ref_audio = ref_audio.mean(0).unsqueeze(0)
+        if ref_sr != 24000:
+            ref_audio = resample(ref_audio, ref_sr, self.configs.device)
+
+        mel2 = mel_fn(ref_audio)
+        mel2 = norm_spec(mel2)
+        T_min = min(mel2.shape[2], fea_ref.shape[2])
+        mel2 = mel2[:, :, :T_min]
+        fea_ref = fea_ref[:, :, :T_min]
+        if T_min > 468:
+            mel2 = mel2[:, :, -468:]
+            fea_ref = fea_ref[:, :, -468:]
+            T_min = 468
+        chunk_len = 934 - T_min
+
+        mel2 = mel2.to(self.precision)
+        fea_todo, ge = self.vits_model.decode_encp(semantic_tokens, phones, refer_audio_spec, ge, speed)
+
+        cfm_resss = []
+        idx = 0
+        while 1:
+            fea_todo_chunk = fea_todo[:, :, idx : idx + chunk_len]
+            if fea_todo_chunk.shape[-1] == 0:
+                break
+            idx += chunk_len
+            fea = torch.cat([fea_ref, fea_todo_chunk], 2).transpose(2, 1)
+
+            cfm_res = self.vits_model.cfm.inference(
+                fea, torch.LongTensor([fea.size(1)]).to(fea.device), mel2, sample_steps, inference_cfg_rate=0
+            )
+            cfm_res = cfm_res[:, :, mel2.shape[2] :]
+
+            mel2 = cfm_res[:, :, -T_min:]
+            fea_ref = fea_todo_chunk[:, :, -T_min:]
+
+            cfm_resss.append(cfm_res)
+        cfm_res = torch.cat(cfm_resss, 2)
+        cfm_res = denorm_spec(cfm_res)
+
+        with torch.inference_mode():
+            wav_gen = self.bigvgan_model(cfm_res)
+            audio = wav_gen[0][0]  # .cpu().detach().numpy()
+
+        return audio
+
+    def v3_synthesis_batched_infer(
+        self,
+        idx_list: List[int],
+        semantic_tokens_list: List[torch.Tensor],
+        batch_phones: List[torch.Tensor],
+        speed: float = 1.0,
+        sample_steps: int = 32,
+    ) -> List[torch.Tensor]:
+        prompt_semantic_tokens = self.prompt_cache["prompt_semantic"].unsqueeze(0).unsqueeze(0).to(self.configs.device)
+        prompt_phones = torch.LongTensor(self.prompt_cache["phones"]).unsqueeze(0).to(self.configs.device)
+        refer_audio_spec = self.prompt_cache["refer_spec"][0].to(dtype=self.precision, device=self.configs.device)
+
+        fea_ref, ge = self.vits_model.decode_encp(prompt_semantic_tokens, prompt_phones, refer_audio_spec)
+        ref_audio: torch.Tensor = self.prompt_cache["raw_audio"]
+        ref_sr = self.prompt_cache["raw_sr"]
+        ref_audio = ref_audio.to(self.configs.device).float()
+        if ref_audio.shape[0] == 2:
+            ref_audio = ref_audio.mean(0).unsqueeze(0)
+        if ref_sr != 24000:
+            ref_audio = resample(ref_audio, ref_sr, self.configs.device)
+
+        mel2 = mel_fn(ref_audio)
+        mel2 = norm_spec(mel2)
+        T_min = min(mel2.shape[2], fea_ref.shape[2])
+        mel2 = mel2[:, :, :T_min]
+        fea_ref = fea_ref[:, :, :T_min]
+        if T_min > 468:
+            mel2 = mel2[:, :, -468:]
+            fea_ref = fea_ref[:, :, -468:]
+            T_min = 468
+        chunk_len = 934 - T_min
+
+        mel2 = mel2.to(self.precision)
+
+        # #### batched inference
+        overlapped_len = 12
+        feat_chunks = []
+        feat_lens = []
+        feat_list = []
+
+        for i, idx in enumerate(idx_list):
+            phones = batch_phones[i].unsqueeze(0).to(self.configs.device)
+            semantic_tokens = (
+                semantic_tokens_list[i][-idx:].unsqueeze(0).unsqueeze(0)
+            )  # .unsqueeze(0)#mq要多unsqueeze一次
+            feat, _ = self.vits_model.decode_encp(semantic_tokens, phones, refer_audio_spec, ge, speed)
+            feat_list.append(feat)
+            feat_lens.append(feat.shape[2])
+
+        feats = torch.cat(feat_list, 2)
+        feats_padded = F.pad(feats, (overlapped_len, 0), "constant", 0)
+        pos = 0
+        padding_len = 0
+        while True:
+            if pos == 0:
+                chunk = feats_padded[:, :, pos : pos + chunk_len]
+            else:
+                pos = pos - overlapped_len
+                chunk = feats_padded[:, :, pos : pos + chunk_len]
+            pos += chunk_len
+            if chunk.shape[-1] == 0:
+                break
+
+            # padding for the last chunk
+            padding_len = chunk_len - chunk.shape[2]
+            if padding_len != 0:
+                chunk = F.pad(chunk, (0, padding_len), "constant", 0)
+            feat_chunks.append(chunk)
+
+        feat_chunks = torch.cat(feat_chunks, 0)
+        bs = feat_chunks.shape[0]
+        fea_ref = fea_ref.repeat(bs, 1, 1)
+        fea = torch.cat([fea_ref, feat_chunks], 2).transpose(2, 1)
+        pred_spec = self.vits_model.cfm.inference(
+            fea, torch.LongTensor([fea.size(1)]).to(fea.device), mel2, sample_steps, inference_cfg_rate=0
+        )
+        pred_spec = pred_spec[:, :, -chunk_len:]
+        dd = pred_spec.shape[1]
+        pred_spec = pred_spec.permute(1, 0, 2).contiguous().view(dd, -1).unsqueeze(0)
+        # pred_spec = pred_spec[..., :-padding_len]
+
+        pred_spec = denorm_spec(pred_spec)
+
+        with torch.no_grad():
+            wav_gen = self.bigvgan_model(pred_spec)
+            audio = wav_gen[0][0]  # .cpu().detach().numpy()
+
+        audio_fragments = []
+        upsample_rate = 256
+        pos = 0
+
+        while pos < audio.shape[-1]:
+            audio_fragment = audio[pos : pos + chunk_len * upsample_rate]
+            audio_fragments.append(audio_fragment)
+            pos += chunk_len * upsample_rate
+
+        audio = self.sola_algorithm(audio_fragments, overlapped_len * upsample_rate)
+        audio = audio[overlapped_len * upsample_rate : -padding_len * upsample_rate]
+
+        audio_fragments = []
+        for feat_len in feat_lens:
+            audio_fragment = audio[: feat_len * upsample_rate]
+            audio_fragments.append(audio_fragment)
+            audio = audio[feat_len * upsample_rate :]
+
+        return audio_fragments
+
+    def sola_algorithm(
+        self,
+        audio_fragments: List[torch.Tensor],
+        overlap_len: int,
+    ):
+        for i in range(len(audio_fragments) - 1):
+            f1 = audio_fragments[i]
+            f2 = audio_fragments[i + 1]
+            w1 = f1[-overlap_len:]
+            w2 = f2[:overlap_len]
+            assert w1.shape == w2.shape
+            corr = F.conv1d(w1.view(1, 1, -1), w2.view(1, 1, -1), padding=w2.shape[-1] // 2).view(-1)[:-1]
+            idx = corr.argmax()
+            f1_ = f1[: -(overlap_len - idx)]
+            audio_fragments[i] = f1_
+
+            f2_ = f2[idx:]
+            window = torch.hann_window((overlap_len - idx) * 2, device=f1.device, dtype=f1.dtype)
+            f2_[: (overlap_len - idx)] = (
+                window[: (overlap_len - idx)] * f2_[: (overlap_len - idx)]
+                + window[(overlap_len - idx) :] * f1[-(overlap_len - idx) :]
+            )
+            audio_fragments[i + 1] = f2_
+
+        return torch.cat(audio_fragments, 0)
diff --git a/GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py b/GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py
new file mode 100644
index 0000000000000000000000000000000000000000..426929f8bdfbde172d584c51810751e3d3635fac
--- /dev/null
+++ b/GPT_SoVITS/TTS_infer_pack/TextPreprocessor.py
@@ -0,0 +1,237 @@
+import os
+import sys
+import threading
+
+from tqdm import tqdm
+
+now_dir = os.getcwd()
+sys.path.append(now_dir)
+
+import re
+import torch
+from text.LangSegmenter import LangSegmenter
+from text import chinese
+from typing import Dict, List, Tuple
+from text.cleaner import clean_text
+from text import cleaned_text_to_sequence
+from transformers import AutoModelForMaskedLM, AutoTokenizer
+from TTS_infer_pack.text_segmentation_method import split_big_text, splits, get_method as get_seg_method
+
+from tools.i18n.i18n import I18nAuto, scan_language_list
+
+language = os.environ.get("language", "Auto")
+language = sys.argv[-1] if sys.argv[-1] in scan_language_list() else language
+i18n = I18nAuto(language=language)
+punctuation = set(["!", "?", "…", ",", ".", "-"])
+
+
+def get_first(text: str) -> str:
+    pattern = "[" + "".join(re.escape(sep) for sep in splits) + "]"
+    text = re.split(pattern, text)[0].strip()
+    return text
+
+
+def merge_short_text_in_array(texts: str, threshold: int) -> list:
+    if (len(texts)) < 2:
+        return texts
+    result = []
+    text = ""
+    for ele in texts:
+        text += ele
+        if len(text) >= threshold:
+            result.append(text)
+            text = ""
+    if len(text) > 0:
+        if len(result) == 0:
+            result.append(text)
+        else:
+            result[len(result) - 1] += text
+    return result
+
+
+class TextPreprocessor:
+    def __init__(self, bert_model: AutoModelForMaskedLM, tokenizer: AutoTokenizer, device: torch.device):
+        self.bert_model = bert_model
+        self.tokenizer = tokenizer
+        self.device = device
+        self.bert_lock = threading.RLock()
+
+    def preprocess(self, text: str, lang: str, text_split_method: str, version: str = "v2") -> List[Dict]:
+        print(f"############ {i18n('切分文本')} ############")
+        text = self.replace_consecutive_punctuation(text)
+        texts = self.pre_seg_text(text, lang, text_split_method)
+        result = []
+        print(f"############ {i18n('提取文本Bert特征')} ############")
+        for text in tqdm(texts):
+            phones, bert_features, norm_text = self.segment_and_extract_feature_for_text(text, lang, version)
+            if phones is None or norm_text == "":
+                continue
+            res = {
+                "phones": phones,
+                "bert_features": bert_features,
+                "norm_text": norm_text,
+            }
+            result.append(res)
+        return result
+
+    def pre_seg_text(self, text: str, lang: str, text_split_method: str):
+        text = text.strip("\n")
+        if len(text) == 0:
+            return []
+        if text[0] not in splits and len(get_first(text)) < 4:
+            text = "。" + text if lang != "en" else "." + text
+        print(i18n("实际输入的目标文本:"))
+        print(text)
+
+        seg_method = get_seg_method(text_split_method)
+        text = seg_method(text)
+
+        while "\n\n" in text:
+            text = text.replace("\n\n", "\n")
+
+        _texts = text.split("\n")
+        _texts = self.filter_text(_texts)
+        _texts = merge_short_text_in_array(_texts, 5)
+        texts = []
+
+        for text in _texts:
+            # 解决输入目标文本的空行导致报错的问题
+            if len(text.strip()) == 0:
+                continue
+            if not re.sub("\W+", "", text):
+                # 检测一下，如果是纯符号，就跳过。
+                continue
+            if text[-1] not in splits:
+                text += "。" if lang != "en" else "."
+
+            # 解决句子过长导致Bert报错的问题
+            if len(text) > 510:
+                texts.extend(split_big_text(text))
+            else:
+                texts.append(text)
+
+        print(i18n("实际输入的目标文本(切句后):"))
+        print(texts)
+        return texts
+
+    def segment_and_extract_feature_for_text(
+        self, text: str, language: str, version: str = "v1"
+    ) -> Tuple[list, torch.Tensor, str]:
+        return self.get_phones_and_bert(text, language, version)
+
+    def get_phones_and_bert(self, text: str, language: str, version: str, final: bool = False):
+        with self.bert_lock:
+            if language in {"en", "all_zh", "all_ja", "all_ko", "all_yue"}:
+                # language = language.replace("all_","")
+                formattext = text
+                while "  " in formattext:
+                    formattext = formattext.replace("  ", " ")
+                if language == "all_zh":
+                    if re.search(r"[A-Za-z]", formattext):
+                        formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
+                        formattext = chinese.mix_text_normalize(formattext)
+                        return self.get_phones_and_bert(formattext, "zh", version)
+                    else:
+                        phones, word2ph, norm_text = self.clean_text_inf(formattext, language, version)
+                        bert = self.get_bert_feature(norm_text, word2ph).to(self.device)
+                elif language == "all_yue" and re.search(r"[A-Za-z]", formattext):
+                    formattext = re.sub(r"[a-z]", lambda x: x.group(0).upper(), formattext)
+                    formattext = chinese.mix_text_normalize(formattext)
+                    return self.get_phones_and_bert(formattext, "yue", version)
+                else:
+                    phones, word2ph, norm_text = self.clean_text_inf(formattext, language, version)
+                    bert = torch.zeros(
+                        (1024, len(phones)),
+                        dtype=torch.float32,
+                    ).to(self.device)
+            elif language in {"zh", "ja", "ko", "yue", "auto", "auto_yue"}:
+                textlist = []
+                langlist = []
+                if language == "auto":
+                    for tmp in LangSegmenter.getTexts(text):
+                        langlist.append(tmp["lang"])
+                        textlist.append(tmp["text"])
+                elif language == "auto_yue":
+                    for tmp in LangSegmenter.getTexts(text):
+                        if tmp["lang"] == "zh":
+                            tmp["lang"] = "yue"
+                        langlist.append(tmp["lang"])
+                        textlist.append(tmp["text"])
+                else:
+                    for tmp in LangSegmenter.getTexts(text):
+                        if tmp["lang"] == "en":
+                            langlist.append(tmp["lang"])
+                        else:
+                            # 因无法区别中日韩文汉字,以用户输入为准
+                            langlist.append(language)
+                        textlist.append(tmp["text"])
+                # print(textlist)
+                # print(langlist)
+                phones_list = []
+                bert_list = []
+                norm_text_list = []
+                for i in range(len(textlist)):
+                    lang = langlist[i]
+                    phones, word2ph, norm_text = self.clean_text_inf(textlist[i], lang, version)
+                    bert = self.get_bert_inf(phones, word2ph, norm_text, lang)
+                    phones_list.append(phones)
+                    norm_text_list.append(norm_text)
+                    bert_list.append(bert)
+                bert = torch.cat(bert_list, dim=1)
+                phones = sum(phones_list, [])
+                norm_text = "".join(norm_text_list)
+
+            if not final and len(phones) < 6:
+                return self.get_phones_and_bert("." + text, language, version, final=True)
+
+            return phones, bert, norm_text
+
+    def get_bert_feature(self, text: str, word2ph: list) -> torch.Tensor:
+        with torch.no_grad():
+            inputs = self.tokenizer(text, return_tensors="pt")
+            for i in inputs:
+                inputs[i] = inputs[i].to(self.device)
+            res = self.bert_model(**inputs, output_hidden_states=True)
+            res = torch.cat(res["hidden_states"][-3:-2], -1)[0].cpu()[1:-1]
+        assert len(word2ph) == len(text)
+        phone_level_feature = []
+        for i in range(len(word2ph)):
+            repeat_feature = res[i].repeat(word2ph[i], 1)
+            phone_level_feature.append(repeat_feature)
+        phone_level_feature = torch.cat(phone_level_feature, dim=0)
+        return phone_level_feature.T
+
+    def clean_text_inf(self, text: str, language: str, version: str = "v2"):
+        language = language.replace("all_", "")
+        phones, word2ph, norm_text = clean_text(text, language, version)
+        phones = cleaned_text_to_sequence(phones, version)
+        return phones, word2ph, norm_text
+
+    def get_bert_inf(self, phones: list, word2ph: list, norm_text: str, language: str):
+        language = language.replace("all_", "")
+        if language == "zh":
+            feature = self.get_bert_feature(norm_text, word2ph).to(self.device)
+        else:
+            feature = torch.zeros(
+                (1024, len(phones)),
+                dtype=torch.float32,
+            ).to(self.device)
+
+        return feature
+
+    def filter_text(self, texts):
+        _text = []
+        if all(text in [None, " ", "\n", ""] for text in texts):
+            raise ValueError(i18n("请输入有效文本"))
+        for text in texts:
+            if text in [None, " ", ""]:
+                pass
+            else:
+                _text.append(text)
+        return _text
+
+    def replace_consecutive_punctuation(self, text):
+        punctuations = "".join(re.escape(p) for p in punctuation)
+        pattern = f"([{punctuations}])([{punctuations}])+"
+        result = re.sub(pattern, r"\1", text)
+        return result
diff --git a/GPT_SoVITS/TTS_infer_pack/__init__.py b/GPT_SoVITS/TTS_infer_pack/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8579a63215aba6c2dc5674c4e8711256c7e904a7
--- /dev/null
+++ b/GPT_SoVITS/TTS_infer_pack/__init__.py
@@ -0,0 +1 @@
+from . import TTS, text_segmentation_method
diff --git a/GPT_SoVITS/TTS_infer_pack/text_segmentation_method.py b/GPT_SoVITS/TTS_infer_pack/text_segmentation_method.py
new file mode 100644
index 0000000000000000000000000000000000000000..fda70a49834ea43c2a3a55154705e111b24fa196
--- /dev/null
+++ b/GPT_SoVITS/TTS_infer_pack/text_segmentation_method.py
@@ -0,0 +1,189 @@
+import re
+from typing import Callable
+
+punctuation = set(["!", "?", "…", ",", ".", "-", " "])
+METHODS = dict()
+
+
+def get_method(name: str) -> Callable:
+    method = METHODS.get(name, None)
+    if method is None:
+        raise ValueError(f"Method {name} not found")
+    return method
+
+
+def get_method_names() -> list:
+    return list(METHODS.keys())
+
+
+def register_method(name):
+    def decorator(func):
+        METHODS[name] = func
+        return func
+
+    return decorator
+
+
+splits = {
+    "，",
+    "。",
+    "？",
+    "！",
+    ",",
+    ".",
+    "?",
+    "!",
+    "~",
+    ":",
+    "：",
+    "—",
+    "…",
+}
+
+
+def split_big_text(text, max_len=510):
+    # 定义全角和半角标点符号
+    punctuation = "".join(splits)
+
+    # 切割文本
+    segments = re.split("([" + punctuation + "])", text)
+
+    # 初始化结果列表和当前片段
+    result = []
+    current_segment = ""
+
+    for segment in segments:
+        # 如果当前片段加上新的片段长度超过max_len，就将当前片段加入结果列表，并重置当前片段
+        if len(current_segment + segment) > max_len:
+            result.append(current_segment)
+            current_segment = segment
+        else:
+            current_segment += segment
+
+    # 将最后一个片段加入结果列表
+    if current_segment:
+        result.append(current_segment)
+
+    return result
+
+
+def split(todo_text):
+    todo_text = todo_text.replace("……", "。").replace("——", "，")
+    if todo_text[-1] not in splits:
+        todo_text += "。"
+    i_split_head = i_split_tail = 0
+    len_text = len(todo_text)
+    todo_texts = []
+    while 1:
+        if i_split_head >= len_text:
+            break  # 结尾一定有标点，所以直接跳出即可，最后一段在上次已加入
+        if todo_text[i_split_head] in splits:
+            i_split_head += 1
+            todo_texts.append(todo_text[i_split_tail:i_split_head])
+            i_split_tail = i_split_head
+        else:
+            i_split_head += 1
+    return todo_texts
+
+
+# 不切
+@register_method("cut0")
+def cut0(inp):
+    if not set(inp).issubset(punctuation):
+        return inp
+    else:
+        return "/n"
+
+
+# 凑四句一切
+@register_method("cut1")
+def cut1(inp):
+    inp = inp.strip("\n")
+    inps = split(inp)
+    split_idx = list(range(0, len(inps), 4))
+    split_idx[-1] = None
+    if len(split_idx) > 1:
+        opts = []
+        for idx in range(len(split_idx) - 1):
+            opts.append("".join(inps[split_idx[idx] : split_idx[idx + 1]]))
+    else:
+        opts = [inp]
+    opts = [item for item in opts if not set(item).issubset(punctuation)]
+    return "\n".join(opts)
+
+
+# 凑50字一切
+@register_method("cut2")
+def cut2(inp):
+    inp = inp.strip("\n")
+    inps = split(inp)
+    if len(inps) < 2:
+        return inp
+    opts = []
+    summ = 0
+    tmp_str = ""
+    for i in range(len(inps)):
+        summ += len(inps[i])
+        tmp_str += inps[i]
+        if summ > 50:
+            summ = 0
+            opts.append(tmp_str)
+            tmp_str = ""
+    if tmp_str != "":
+        opts.append(tmp_str)
+    # print(opts)
+    if len(opts) > 1 and len(opts[-1]) < 50:  ##如果最后一个太短了，和前一个合一起
+        opts[-2] = opts[-2] + opts[-1]
+        opts = opts[:-1]
+    opts = [item for item in opts if not set(item).issubset(punctuation)]
+    return "\n".join(opts)
+
+
+# 按中文句号。切
+@register_method("cut3")
+def cut3(inp):
+    inp = inp.strip("\n")
+    opts = ["%s" % item for item in inp.strip("。").split("。")]
+    opts = [item for item in opts if not set(item).issubset(punctuation)]
+    return "\n".join(opts)
+
+
+# 按英文句号.切
+@register_method("cut4")
+def cut4(inp):
+    inp = inp.strip("\n")
+    opts = re.split(r"(?<!\d)\.(?!\d)", inp.strip("."))
+    opts = [item for item in opts if not set(item).issubset(punctuation)]
+    return "\n".join(opts)
+
+
+# 按标点符号切
+# contributed by https://github.com/AI-Hobbyist/GPT-SoVITS/blob/main/GPT_SoVITS/inference_webui.py
+@register_method("cut5")
+def cut5(inp):
+    inp = inp.strip("\n")
+    punds = {",", ".", ";", "?", "!", "、", "，", "。", "？", "！", ";", "：", "…"}
+    mergeitems = []
+    items = []
+
+    for i, char in enumerate(inp):
+        if char in punds:
+            if char == "." and i > 0 and i < len(inp) - 1 and inp[i - 1].isdigit() and inp[i + 1].isdigit():
+                items.append(char)
+            else:
+                items.append(char)
+                mergeitems.append("".join(items))
+                items = []
+        else:
+            items.append(char)
+
+    if items:
+        mergeitems.append("".join(items))
+
+    opt = [item for item in mergeitems if not set(item).issubset(punds)]
+    return "\n".join(opt)
+
+
+if __name__ == "__main__":
+    method = get_method("cut5")
+    print(method("你好，我是小明。你好，我是小红。你好，我是小刚。你好，我是小张。"))
diff --git a/GPT_SoVITS/pretrained_models/chinese-hubert-base/pytorch_model.bin b/GPT_SoVITS/pretrained_models/chinese-hubert-base/pytorch_model.bin
new file mode 100644
index 0000000000000000000000000000000000000000..a73dedefbc10383bc70e289fe0897d2157c6f66b
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/chinese-hubert-base/pytorch_model.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24164f129c66499d1346e2aa55f183250c223161ec2770c0da3d3b08cf432d3c
+size 188811417
diff --git a/GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large/pytorch_model.bin b/GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large/pytorch_model.bin
new file mode 100644
index 0000000000000000000000000000000000000000..9d67762821a81a27af2c60d8094ad86cad6510c3
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large/pytorch_model.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e53a693acc59ace251d143d068096ae0d7b79e4b1b503fa84c9dcf576448c1d8
+size 651225145
diff --git a/GPT_SoVITS/pretrained_models/fast_langdetect/lid.176.bin b/GPT_SoVITS/pretrained_models/fast_langdetect/lid.176.bin
new file mode 100644
index 0000000000000000000000000000000000000000..f8707035ea3cc86ac248a4e31fa6368cd845476a
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/fast_langdetect/lid.176.bin
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e69ec5451bc261cc7844e49e4792a85d7f09c06789ec800fc4a44aec362764e
+size 131266198
diff --git a/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt
new file mode 100644
index 0000000000000000000000000000000000000000..3b8a968297b05d69f441e5981111d48426fdcb9b
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s1bert25hz-5kh-longer-epoch=12-step=369668.ckpt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:732f94e63b148066e24c7f9d2637f3374083e637635f07fbdb695dee20ddbe1f
+size 155315150
diff --git a/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2D2333k.pth b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2D2333k.pth
new file mode 100644
index 0000000000000000000000000000000000000000..5eecfd7267624bfb64cd8e3be88f1e1d2d99e19b
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2D2333k.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ae7fe8dd8c8f2e718de359e00edac88b0c71ab2fd10b07ad4cc45070eb8a836
+size 93534164
diff --git a/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth
new file mode 100644
index 0000000000000000000000000000000000000000..51d3419ff8403b6d73334b4dba7aabdd79ed38fe
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/gsv-v2final-pretrained/s2G2333k.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:924fdccaa3c574bf139c25c9759aa1ed3b3f99e19a7c529ee996c2bc17663695
+size 106035259
diff --git a/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth b/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth
new file mode 100644
index 0000000000000000000000000000000000000000..0176922b6a3ddd823b3357fd934c14279980c583
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/s2Gv4.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:906fe22f48c3e037a389df291d4d32a9414e15dbb8f9628643e83aaced109ea4
+size 769025545
diff --git a/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth b/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth
new file mode 100644
index 0000000000000000000000000000000000000000..6691a922691bb7fab617b5aed395b1566e4cec87
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/gsv-v4-pretrained/vocoder.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d611913df7b12d49e8976c944558d2d096816365edfc6c35a9e85b67dd14ed9
+size 57781109
diff --git a/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt b/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt
new file mode 100644
index 0000000000000000000000000000000000000000..a0fe198f23ef09727e6b0dfdebd5ea14e1d16a9f
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/models--nvidia--bigvgan_v2_24khz_100band_256x/bigvgan_generator.pt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee5e2f9cd60b51db75e1806f4fe7621733757586c541c78cb3dd369d5ba24476
+size 225179685
diff --git a/GPT_SoVITS/pretrained_models/s1v3.ckpt b/GPT_SoVITS/pretrained_models/s1v3.ckpt
new file mode 100644
index 0000000000000000000000000000000000000000..54a87a3557ca48356e31cf3651155a6dbb8a05da
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/s1v3.ckpt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:87133414860ea14ff6620c483a3db5ed07b44be42e2c3fcdad65523a729a745a
+size 155284856
diff --git a/GPT_SoVITS/pretrained_models/s2G488k.pth b/GPT_SoVITS/pretrained_models/s2G488k.pth
new file mode 100644
index 0000000000000000000000000000000000000000..a63e37a8755ad141ba68db9ef0109bf36ec020c4
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/s2G488k.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:020a014e1e01e550e510f2f61fae5e5f5b6aab40f15c22f1f12f724df507e835
+size 105973721
diff --git a/GPT_SoVITS/pretrained_models/s2Gv3.pth b/GPT_SoVITS/pretrained_models/s2Gv3.pth
new file mode 100644
index 0000000000000000000000000000000000000000..e0f4efe7cf733b418fe16663d7c2fd90abd5a2bc
--- /dev/null
+++ b/GPT_SoVITS/pretrained_models/s2Gv3.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f33abb1920076d988e1711d5f41b5c9c6d7f92575b4acf0ad4fae6a4ebf0cf19
+size 769035145
diff --git a/GPT_SoVITS/text/G2PWModel/g2pW.onnx b/GPT_SoVITS/text/G2PWModel/g2pW.onnx
new file mode 100644
index 0000000000000000000000000000000000000000..0fbd9f7affb0fb08156a85be11ef7ba3602f5f44
--- /dev/null
+++ b/GPT_SoVITS/text/G2PWModel/g2pW.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2eb3c71fd95117b2e1abef8d2d0cd78aae894bbe7f0fac105ddc9c32ce63cbd0
+size 635212732
diff --git a/GPT_SoVITS/text/engdict_cache.pickle b/GPT_SoVITS/text/engdict_cache.pickle
new file mode 100644
index 0000000000000000000000000000000000000000..cdca61190dbe75c37ce3a05e8d5c942fc1a2d726
--- /dev/null
+++ b/GPT_SoVITS/text/engdict_cache.pickle
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bff9393f4b192d873a11335efc8f124771087b6dc847d34fd240c2846889d2b
+size 5965909
diff --git a/GPT_SoVITS/text/g2pw/polyphonic.pickle b/GPT_SoVITS/text/g2pw/polyphonic.pickle
new file mode 100644
index 0000000000000000000000000000000000000000..0a8912d1b2cb3016eac46dc55ec6e10a7b442f7b
--- /dev/null
+++ b/GPT_SoVITS/text/g2pw/polyphonic.pickle
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f425246160a32c578557cd3151cd0bb97f5f44c3aaf65e718dd2c3213c04fb4b
+size 1322387
diff --git a/GPT_SoVITS/text/ja_userdic/userdict.csv b/GPT_SoVITS/text/ja_userdic/userdict.csv
new file mode 100644
index 0000000000000000000000000000000000000000..d9aa2378c16a27eaa512a2a7642d7241bd050e8b
--- /dev/null
+++ b/GPT_SoVITS/text/ja_userdic/userdict.csv
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d857e443ee48d9641096816a98996669602895411e4330d7d91d1dbe1103389f
+size 17180971
diff --git a/GPT_SoVITS/text/namedict_cache.pickle b/GPT_SoVITS/text/namedict_cache.pickle
new file mode 100644
index 0000000000000000000000000000000000000000..9ad2ed9a6478077f8fe4469192597fa8f6442cf6
--- /dev/null
+++ b/GPT_SoVITS/text/namedict_cache.pickle
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:559552094c4a6e995213e3fa586330e078ef8cb3a7a95a3109e945111cd2bfc1
+size 760663
diff --git a/docs/cn/Changelog_CN.md b/docs/cn/Changelog_CN.md
new file mode 100644
index 0000000000000000000000000000000000000000..cd1d50d0ef4caec723f651c84f616ae8cc0b8f50
--- /dev/null
+++ b/docs/cn/Changelog_CN.md
@@ -0,0 +1,302 @@
+### 20240121更新
+
+1-config添加is_share, 诸如colab等场景可以将此改为True, 来使得webui映射到公网
+
+2-WebUI添加英文系统英文翻译适配
+
+3-cmd-asr自动判断是否已自带damo模型, 如不在默认目录上将从modelscope自带下载
+
+4-[SoVITS训练报错ZeroDivisionError](https://github.com/RVC-Boss/GPT-SoVITS/issues/79) 尝试修复(过滤长度0的样本等)
+
+5-清理TEMP文件夹缓存音频等文件
+
+6-大幅削弱合成音频包含参考音频结尾的问题
+
+### 20240122更新
+
+1-修复过短输出文件返回重复参考音频的问题.
+
+2-经测试, 英文日文训练原生支持(日文训练需要根目录不含非英文等特殊字符).
+
+3-音频路径检查.如果尝试读取输入错的路径报错路径不存在, 而非ffmpeg错误.
+
+### 20240123更新
+
+1-解决hubert提取nan导致SoVITS/GPT训练报错ZeroDivisionError的问题
+
+2-支持推理界面快速切换模型
+
+3-优化模型文件排序逻辑
+
+4-中文分词使用jieba_fast代替jieba
+
+### 20240126更新
+
+1-支持输出文本中英混合、日英混合
+
+2-输出可选切分模式
+
+3-修复uvr5读取到目录自动跳出的问题
+
+4-修复多个换行导致推理报错
+
+5-去除推理界面大量冗余log
+
+6-支持mac训练推理
+
+7-自动识别不支持半精度的卡强制单精度.cpu推理下强制单精度.
+
+### 20240128更新
+
+1-修复数字转汉字念法问题
+
+2-修复句首少量字容易吞字的问题
+
+3-通过限制排除不合理的参考音频长度
+
+4-修复GPT训练不保存ckpt的问题
+
+5-完善Dockerfile的下载模型流程
+
+### 20240129更新
+
+1-16系等半精度训练有问题的显卡把训练配置改为单精度训练
+
+2-测试更新可用的colab版本
+
+3-修复git clone modelscope funasr仓库+老版本funasr导致接口不对齐报错的问题
+
+
+### 20240130更新
+
+1-所有涉及路径的地方双引号自动去除,小白复制路径带双引号不会报错
+
+2-修复中英文标点切割问题和句首句尾补标点的问题
+
+3-增加按标点符号切分
+
+### 20240201更新
+
+1-修复uvr5读取格式错误导致分离失败的问题
+
+2-支持中日英混合多种文本自动切分识别语种
+
+### 20240202更新
+
+1-修复asr路径尾缀带/保存文件名报错
+
+2-引入paddlespeech的Normalizer https://github.com/RVC-Boss/GPT-SoVITS/pull/377 修复一些问题, 例如: xx.xx%(带百分号类), 元/吨 会读成 元吨 而不是元每吨,下划线不再会报错
+
+### 20240207更新
+
+1-修正语种传参混乱导致中文推理效果下降 https://github.com/RVC-Boss/GPT-SoVITS/issues/391
+
+2-uvr5适配高版本librosa https://github.com/RVC-Boss/GPT-SoVITS/pull/403
+
+3-[修复uvr5 inf everywhere报错的问题(is_half传参未转换bool导致恒定半精度推理, 16系显卡会inf)](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8)
+
+4-优化英文文本前端
+
+5-修复gradio依赖
+
+6-支持三连根目录留空自动读取.list全路径
+
+7-集成faster whisper ASR日文英文
+
+### 20240208更新
+
+1-GPT训练卡死 (win10 1909) 和https://github.com/RVC-Boss/GPT-SoVITS/issues/232 (系统语言繁体) GPT训练报错, [尝试修复](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b).
+
+### 20240212更新
+
+1-faster whisper和funasr逻辑优化.faster whisper转镜像站下载, 规避huggingface连不上的问题.
+
+2-DPO Loss实验性训练选项开启, 通过构造负样本训练缓解GPT重复漏字问题.推理界面公开几个推理参数. https://github.com/RVC-Boss/GPT-SoVITS/pull/457
+
+### 20240214更新
+
+1-训练支持中文实验名 (原来会报错)
+
+2-DPO训练改为可勾选选项而非必须.如勾选batch size自动减半.修复推理界面新参数不传参的问题.
+
+### 20240216更新
+
+1-支持无参考文本输入
+
+2-修复中文文本前端bug https://github.com/RVC-Boss/GPT-SoVITS/issues/475
+
+### 20240221更新
+
+1-数据处理添加语音降噪选项 (降噪为只剩16k采样率, 除非底噪很大先不急着用哦).
+
+2-中文日文前端处理优化 https://github.com/RVC-Boss/GPT-SoVITS/pull/559 https://github.com/RVC-Boss/GPT-SoVITS/pull/556 https://github.com/RVC-Boss/GPT-SoVITS/pull/532 https://github.com/RVC-Boss/GPT-SoVITS/pull/507 https://github.com/RVC-Boss/GPT-SoVITS/pull/509
+
+3-mac CPU推理更快因此把推理设备从mps改到CPU
+
+4-colab修复不开启公网url
+
+### 20240306更新
+
+1-推理加速50% (RTX3090+pytorch2.2.1+cu11.8+win10+py39 tested) https://github.com/RVC-Boss/GPT-SoVITS/pull/672
+
+2-如果用faster whisper非中文ASR不再需要先下中文funasr模型
+
+3-修复uvr5去混响模型 是否混响 反的 https://github.com/RVC-Boss/GPT-SoVITS/pull/610
+
+4-faster whisper如果无cuda可用自动cpu推理 https://github.com/RVC-Boss/GPT-SoVITS/pull/675
+
+5-修改is_half的判断使在Mac上能正常CPU推理 https://github.com/RVC-Boss/GPT-SoVITS/pull/573
+
+### 202403/202404/202405更新
+
+2个重点
+
+1-修复sovits训练未冻结vq的问题 (可能造成效果下降)
+
+2-增加一个快速推理分支
+
+以下都是小修补
+
+1-修复无参考文本模式问题
+
+2-优化中英文文本前端
+
+3-api格式优化
+
+4-cmd格式问题修复
+
+5-训练数据处理阶段不支持的语言提示报错
+
+6-nan自动转fp32阶段的hubert提取bug修复
+
+### 20240610
+
+小问题修复: 
+
+1-完善纯标点、多标点文本输入的判断逻辑 https://github.com/RVC-Boss/GPT-SoVITS/pull/1168 https://github.com/RVC-Boss/GPT-SoVITS/pull/1169
+
+2-uvr5中的mdxnet去混响cmd格式修复, 兼容路径带空格  [#501a74a](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232)
+
+3-s2训练进度条逻辑修复 https://github.com/RVC-Boss/GPT-SoVITS/pull/1159
+
+大问题修复: 
+
+4-修复了webui的GPT中文微调没读到bert导致和推理不一致, 训练太多可能效果还会变差的问题.如果大量数据微调的建议重新微调模型得到质量优化 [#99f09c8](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a)
+
+### 20240706
+
+小问题修复: 
+
+1-[修正CPU推理默认bs小数](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041)
+
+2-修复降噪、asr中途遇到异常跳出所有需处理的音频文件的问题 https://github.com/RVC-Boss/GPT-SoVITS/pull/1258 https://github.com/RVC-Boss/GPT-SoVITS/pull/1265 https://github.com/RVC-Boss/GPT-SoVITS/pull/1267
+
+3-修复按标点符号切分时小数会被切分 https://github.com/RVC-Boss/GPT-SoVITS/pull/1253
+
+4-[多卡训练多进程保存逻辑修复](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca)
+
+5-移除冗余my_utils https://github.com/RVC-Boss/GPT-SoVITS/pull/1251
+
+重点: 
+
+6-倍速推理代码经过验证后推理效果和base完全一致, 合并进main.使用的代码: https://github.com/RVC-Boss/GPT-SoVITS/pull/672 .支持无参考文本模式也倍速.
+
+后面会逐渐验证快速推理分支的推理改动的一致性
+
+### 20240727
+
+1-清理冗余i18n代码 https://github.com/RVC-Boss/GPT-SoVITS/pull/1298
+
+2-修复用户打文件及路径在结尾添加/会导致命令行报错的问题 https://github.com/RVC-Boss/GPT-SoVITS/pull/1299
+
+3-修复GPT训练的step计算逻辑 https://github.com/RVC-Boss/GPT-SoVITS/pull/756
+
+重点: 
+
+4-[支持合成语速调节.支持冻结随机性只调节语速, ](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2)并将其更新到api.py上https://github.com/RVC-Boss/GPT-SoVITS/pull/1340
+
+
+### 20240806
+
+1-增加bs-roformer人声伴奏分离模型支持. https://github.com/RVC-Boss/GPT-SoVITS/pull/1306 https://github.com/RVC-Boss/GPT-SoVITS/pull/1356 [支持fp16推理.](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c)
+
+2-更好的中文文本前端. https://github.com/RVC-Boss/GPT-SoVITS/pull/987 https://github.com/RVC-Boss/GPT-SoVITS/pull/1351 https://github.com/RVC-Boss/GPT-SoVITS/pull/1404 优化多音字逻辑 (v2版本特供). https://github.com/RVC-Boss/GPT-SoVITS/pull/488
+
+3-自动填充下一步的文件路径 https://github.com/RVC-Boss/GPT-SoVITS/pull/1355
+
+4-增加喂饭逻辑, 用户瞎写显卡序号也可以正常运作 [bce451a](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299) [4c8b761](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78)
+
+5-增加粤语ASR支持 [8a10147](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3)
+
+6-GPT-SoVITS-v2支持
+
+7-计时逻辑优化 https://github.com/RVC-Boss/GPT-SoVITS/pull/1387
+
+### 20240821
+
+1-fast_inference分支合并进main: https://github.com/RVC-Boss/GPT-SoVITS/pull/1490
+
+2-支持通过ssml标签优化数字、电话、时间日期等: https://github.com/RVC-Boss/GPT-SoVITS/issues/1508
+
+3-api修复优化: https://github.com/RVC-Boss/GPT-SoVITS/pull/1503
+
+4-修复了参考音频混合只能上传一条的bug:https://github.com/RVC-Boss/GPT-SoVITS/pull/1422
+
+5-增加了各种数据集检查,若缺失会弹出warning:https://github.com/RVC-Boss/GPT-SoVITS/pull/1422
+
+### 20250211
+
+增加gpt-sovits-v3模型, 需要14G显存可以微调
+
+### 20250212
+
+sovits-v3微调支持开启梯度检查点, 需要12G显存可以微调https://github.com/RVC-Boss/GPT-SoVITS/pull/2040
+
+### 20250214
+
+优化多语种混合文本切分策略a https://github.com/RVC-Boss/GPT-SoVITS/pull/2047
+
+### 20250217
+
+优化文本里的数字和英文处理逻辑https://github.com/RVC-Boss/GPT-SoVITS/pull/2062
+
+### 20250218
+
+优化多语种混合文本切分策略b https://github.com/RVC-Boss/GPT-SoVITS/pull/2073
+
+### 20250223
+
+1-sovits-v3微调支持lora训练, 需要8G显存可以微调, 效果比全参微调更好
+
+2-人声背景音分离增加mel band roformer模型支持https://github.com/RVC-Boss/GPT-SoVITS/pull/2078
+
+### 20250226
+
+https://github.com/RVC-Boss/GPT-SoVITS/pull/2112 https://github.com/RVC-Boss/GPT-SoVITS/pull/2114
+
+修复中文路径下mecab的报错 (具体表现为日文韩文、文本混合语种切分可能会遇到的报错)
+
+### 20250227
+
+针对v3生成24k音频感觉闷的问题https://github.com/RVC-Boss/GPT-SoVITS/issues/2085 https://github.com/RVC-Boss/GPT-SoVITS/issues/2117 ,支持使用24k to 48k的音频超分模型缓解.
+
+
+### 20250228
+
+修复短文本语种选择出错 https://github.com/RVC-Boss/GPT-SoVITS/pull/2122
+
+修复v3sovits未传参以支持调节语速
+
+### 202503
+
+修复一批由依赖的库版本不对导致的问题https://github.com/RVC-Boss/GPT-SoVITS/commit/6c468583c5566e5fbb4fb805e4cc89c403e997b8
+
+修复模型加载异步逻辑https://github.com/RVC-Boss/GPT-SoVITS/commit/03b662a769946b7a6a8569a354860e8eeeb743aa
+
+修复其他若干bug
+
+重点更新: 
+
+1-v3支持并行推理 https://github.com/RVC-Boss/GPT-SoVITS/commit/03b662a769946b7a6a8569a354860e8eeeb743aa
+
+2-整合包修复onnxruntime GPU推理的支持, 影响:  (1) g2pw有个onnx模型原先是CPU推理现在用GPU, 显著降低推理的CPU瓶颈 (2) foxjoy去混响模型现在可使用GPU推理
diff --git a/docs/cn/README.md b/docs/cn/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..cc72b893eef3f44fb6e564e354dc0e66ba623fb4
--- /dev/null
+++ b/docs/cn/README.md
@@ -0,0 +1,395 @@
+<div align="center">
+
+<h1>GPT-SoVITS-WebUI</h1>
+强大的少样本语音转换与语音合成Web用户界面.<br><br>
+
+[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
+
+<a href="https://trendshift.io/repositories/7033" target="_blank"><img src="https://trendshift.io/api/badge/repositories/7033" alt="RVC-Boss%2FGPT-SoVITS | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+
+<!-- img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br> -->
+
+[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
+[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
+[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
+[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
+
+[**English**](../../README.md) | **中文简体** | [**日本語**](../ja/README.md) | [**한국어**](../ko/README.md) | [**Türkçe**](../tr/README.md)
+
+</div>
+
+---
+
+## 功能: 
+
+1. **零样本文本到语音 (TTS): ** 输入 5 秒的声音样本, 即刻体验文本到语音转换.
+
+2. **少样本 TTS: ** 仅需 1 分钟的训练数据即可微调模型, 提升声音相似度和真实感.
+
+3. **跨语言支持: ** 支持与训练数据集不同语言的推理, 目前支持英语、日语、韩语、粤语和中文.
+
+4. **WebUI 工具: ** 集成工具包括声音伴奏分离、自动训练集分割、中文自动语音识别(ASR)和文本标注, 协助初学者创建训练数据集和 GPT/SoVITS 模型.
+
+**查看我们的介绍视频 [demo video](https://www.bilibili.com/video/BV12g4y1m7Uw)**
+
+未见过的说话者 few-shot 微调演示: 
+
+https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
+
+**用户手册: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
+
+## 安装
+
+中国地区的用户可[点击此处](https://www.codewithgpu.com/i/RVC-Boss/GPT-SoVITS/GPT-SoVITS-Official)使用 AutoDL 云端镜像进行体验.
+
+### 测试通过的环境
+
+| Python Version | PyTorch Version  | Device          |
+|----------------|------------------|-----------------|
+| Python 3.9     | PyTorch 2.0.1    | CUDA 11.8       |
+| Python 3.10.13 | PyTorch 2.1.2    | CUDA 12.3       |
+| Python 3.10.17 | PyTorch 2.5.1    | CUDA 12.4       |
+| Python 3.9     | PyTorch 2.5.1    | Apple silicon   |
+| Python 3.11    | PyTorch 2.6.0    | Apple silicon   |
+| Python 3.9     | PyTorch 2.2.2    | CPU             |
+
+### Windows
+
+如果你是 Windows 用户 (已在 win>=10 上测试), 可以下载[整合包](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-v3lora-20250228.7z?download=true), 解压后双击 go-webui.bat 即可启动 GPT-SoVITS-WebUI.
+
+**中国地区的用户可以[在此处下载整合包](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e/dkxgpiy9zb96hob4#KTvnO).**
+
+### Linux
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### macOS
+
+**注: 在 Mac 上使用 GPU 训练的模型效果显著低于其他设备训练的模型, 所以我们暂时使用 CPU 进行训练.**
+
+1. 运行 `xcode-select --install` 安装 Xcode command-line tools.
+2. 运行以下的命令来安装本项目: 
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### 手动安装
+
+#### 安装 FFmpeg
+
+##### Conda 用户
+
+```bash
+conda install ffmpeg
+```
+
+##### Ubuntu/Debian 用户
+
+```bash
+sudo apt install ffmpeg
+sudo apt install libsox-dev
+conda install -c conda-forge 'ffmpeg<7'
+```
+
+##### Windows 用户
+
+下载并将 [ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) 和 [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) 放置在 GPT-SoVITS 根目录下.
+
+安装 [Visual Studio 2017](https://aka.ms/vs/17/release/vc_redist.x86.exe) 环境(仅限韩语 TTS)
+
+##### MacOS 用户
+
+```bash
+brew install ffmpeg
+```
+
+#### 安装依赖
+
+```bash
+pip install -r extra-req.txt --no-deps
+pip install -r requirements.txt
+```
+
+### 在 Docker 中使用
+
+#### docker-compose.yaml 设置
+
+0. image 的标签: 由于代码库更新很快, 镜像的打包和测试又很慢, 所以请自行在 [Docker Hub](https://hub.docker.com/r/breakstring/gpt-sovits)(旧版本) 查看当前打包好的最新的镜像并根据自己的情况选用, 或者在本地根据您自己的需求通过 Dockerfile 进行构建.
+1. 环境变量: 
+
+- is_half: 半精度/双精度控制.在进行 "SSL extracting" 步骤时如果无法正确生成 4-cnhubert/5-wav32k 目录下的内容时, 一般都是它引起的, 可以根据实际情况来调整为 True 或者 False.
+
+2. Volume 设置, 容器内的应用根目录设置为 /workspace. 默认的 docker-compose.yaml 中列出了一些实际的例子, 便于上传/下载内容.
+3. shm_size: Windows 下的 Docker Desktop 默认可用内存过小, 会导致运行异常, 根据自己情况酌情设置.
+4. deploy 小节下的 gpu 相关内容, 请根据您的系统和实际情况酌情设置.
+
+#### 通过 docker compose 运行
+
+```
+docker compose -f "docker-compose.yaml" up -d
+```
+
+#### 通过 docker 命令运行
+
+同上, 根据您自己的实际情况修改对应的参数, 然后运行如下命令: 
+
+```
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+```
+
+## 预训练模型
+
+**若成功运行`install.sh`可跳过 No.1,2,3**
+
+**中国地区的用户可以[在此处下载这些模型](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e/dkxgpiy9zb96hob4#nVNhX).**
+
+1. 从 [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) 下载预训练模型, 并将其放置在 `GPT_SoVITS/pretrained_models` 目录中.
+
+2. 从 [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) 下载模型, 解压并重命名为 `G2PWModel`, 然后将其放置在 `GPT_SoVITS/text` 目录中. (仅限中文 TTS)
+
+3. 对于 UVR5 (人声/伴奏分离和混响移除, 额外功能), 从 [UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) 下载模型, 并将其放置在 `tools/uvr5/uvr5_weights` 目录中.
+
+   - 如果你在 UVR5 中使用 `bs_roformer` 或 `mel_band_roformer`模型, 你可以手动下载模型和相应的配置文件, 并将它们放在 `tools/UVR5/UVR5_weights` 中.**重命名模型文件和配置文件, 确保除后缀外**, 模型和配置文件具有相同且对应的名称.此外, 模型和配置文件名**必须包含"roformer"**, 才能被识别为 roformer 类的模型.
+
+   - 建议在模型名称和配置文件名中**直接指定模型类型**, 例如`mel_mand_roformer`、`bs_roformer`.如果未指定, 将从配置文中比对特征, 以确定它是哪种类型的模型.例如, 模型`bs_roformer_ep_368_sdr_12.9628.ckpt` 和对应的配置文件`bs_roformer_ep_368_sdr_12.9628.yaml` 是一对.`kim_mel_band_roformer.ckpt` 和 `kim_mel_band_roformer.yaml` 也是一对.
+
+4. 对于中文 ASR (额外功能), 从 [Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files)、[Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files) 和 [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) 下载模型, 并将它们放置在 `tools/asr/models` 目录中.
+
+5. 对于英语或日语 ASR (额外功能), 从 [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) 下载模型, 并将其放置在 `tools/asr/models` 目录中.此外, [其他模型](https://huggingface.co/Systran) 可能具有类似效果且占用更少的磁盘空间.
+
+## 数据集格式
+
+文本到语音 (TTS) 注释 .list 文件格式: 
+
+```
+vocal_path|speaker_name|language|text
+```
+
+语言字典: 
+
+- 'zh': 中文
+- 'ja': 日语
+- 'en': 英语
+- 'ko': 韩语
+- 'yue': 粤语
+
+示例: 
+
+```
+D:\GPT-SoVITS\xxx/xxx.wav|xxx|zh|我爱玩原神.
+```
+
+## 微调与推理
+
+### 打开 WebUI
+
+#### 整合包用户
+
+双击`go-webui.bat`或者使用`go-webui.ps1`
+若想使用 V1,则双击`go-webui-v1.bat`或者使用`go-webui-v1.ps1`
+
+#### 其他
+
+```bash
+python webui.py <language(optional)>
+```
+
+若想使用 V1,则
+
+```bash
+python webui.py v1 <language(optional)>
+```
+
+或者在 webUI 内动态切换
+
+### 微调
+
+#### 现已支持自动填充路径
+
+    1. 填入训练音频路径
+    2. 切割音频
+    3. 进行降噪(可选)
+    4. 进行ASR
+    5. 校对标注
+    6. 前往下一个窗口,点击训练
+
+### 打开推理 WebUI
+
+#### 整合包用户
+
+双击 `go-webui.bat` 或者使用 `go-webui.ps1` ,然后在 `1-GPT-SoVITS-TTS/1C-推理` 中打开推理 webUI
+
+#### 其他
+
+```bash
+python GPT_SoVITS/inference_webui.py <language(optional)>
+```
+
+或者
+
+```bash
+python webui.py
+```
+
+然后在 `1-GPT-SoVITS-TTS/1C-推理` 中打开推理 webUI
+
+## V2 发布说明
+
+新特性:
+
+1. 支持韩语及粤语
+
+2. 更好的文本前端
+
+3. 底模由 2k 小时扩展至 5k 小时
+
+4. 对低音质参考音频 (尤其是来源于网络的高频严重缺失、听着很闷的音频) 合成出来音质更好
+
+   详见[wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v2%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+从 v1 环境迁移至 v2
+
+1. 需要 pip 安装 requirements.txt 更新环境
+
+2. 需要克隆 github 上的最新代码
+
+3. 需要从[huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained) 下载预训练模型文件放到 GPT_SoVITS\pretrained_models\gsv-v2final-pretrained 下
+
+   中文额外需要下载[G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (下载 G2PW 模型,解压并重命名为`G2PWModel`,将其放到`GPT_SoVITS/text`目录下)
+
+## V3 更新说明
+
+新模型特点:
+
+1. 音色相似度更像, 需要更少训练集来逼近本人 (不训练直接使用底模模式下音色相似性提升更大)
+
+2. GPT 合成更稳定, 重复漏字更少, 也更容易跑出丰富情感
+
+   详见[wiki](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v2%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+从 v2 环境迁移至 v3
+
+1. 需要 pip 安装 requirements.txt 更新环境
+
+2. 需要克隆 github 上的最新代码
+
+3. 从[huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main)下载这些 v3 新增预训练模型 (s1v3.ckpt, s2Gv3.pth and models--nvidia--bigvgan_v2_24khz_100band_256x folder)将他们放到`GPT_SoVITS\pretrained_models`目录下
+
+   如果想用音频超分功能缓解 v3 模型生成 24k 音频觉得闷的问题, 需要下载额外的模型参数, 参考[how to download](../../tools/AP_BWE_main/24kto48k/readme.txt)
+
+## 待办事项清单
+
+- [x] **高优先级: **
+
+  - [x] 日语和英语的本地化.
+  - [x] 用户指南.
+  - [x] 日语和英语数据集微调训练.
+
+- [ ] **功能:**
+  - [x] 零样本声音转换 (5 秒) / 少样本声音转换 (1 分钟).
+  - [x] TTS 语速控制.
+  - [ ] ~~增强的 TTS 情感控制.~~
+  - [ ] 尝试将 SoVITS 令牌输入更改为词汇的概率分布.
+  - [x] 改进英语和日语文本前端.
+  - [ ] 开发体积小和更大的 TTS 模型.
+  - [x] Colab 脚本.
+  - [x] 扩展训练数据集 (从 2k 小时到 10k 小时).
+  - [x] 更好的 sovits 基础模型 (增强的音频质量).
+  - [ ] 模型混合.
+
+##  (附加) 命令行运行方式
+
+使用命令行打开 UVR5 的 WebUI
+
+```
+python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
+```
+
+<!-- 如果打不开浏览器, 请按照下面的格式进行UVR处理, 这是使用mdxnet进行音频处理的方式
+````
+python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
+```` -->
+
+这是使用命令行完成数据集的音频切分的方式
+
+```
+python audio_slicer.py \
+    --input_path "<path_to_original_audio_file_or_directory>" \
+    --output_root "<directory_where_subdivided_audio_clips_will_be_saved>" \
+    --threshold <volume_threshold> \
+    --min_length <minimum_duration_of_each_subclip> \
+    --min_interval <shortest_time_gap_between_adjacent_subclips>
+    --hop_size <step_size_for_computing_volume_curve>
+```
+
+这是使用命令行完成数据集 ASR 处理的方式 (仅限中文)
+
+```
+python tools/asr/funasr_asr.py -i <input> -o <output>
+```
+
+通过 Faster_Whisper 进行 ASR 处理 (除中文之外的 ASR 标记)
+
+ (没有进度条, GPU 性能可能会导致时间延迟)
+
+```
+python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p <precision>
+```
+
+启用自定义列表保存路径
+
+## 致谢
+
+特别感谢以下项目和贡献者: 
+
+### 理论研究
+
+- [ar-vits](https://github.com/innnky/ar-vits)
+- [SoundStorm](https://github.com/yangdongchao/SoundStorm/tree/master/soundstorm/s1/AR)
+- [vits](https://github.com/jaywalnut310/vits)
+- [TransferTTS](https://github.com/hcy71o/TransferTTS/blob/master/models.py#L556)
+- [contentvec](https://github.com/auspicious3000/contentvec/)
+- [hifi-gan](https://github.com/jik876/hifi-gan)
+- [fish-speech](https://github.com/fishaudio/fish-speech/blob/main/tools/llama/generate.py#L41)
+- [f5-TTS](https://github.com/SWivid/F5-TTS/blob/main/src/f5_tts/model/backbones/dit.py)
+- [shortcut flow matching](https://github.com/kvfrans/shortcut-models/blob/main/targets_shortcut.py)
+
+### 预训练模型
+
+- [Chinese Speech Pretrain](https://github.com/TencentGameMate/chinese_speech_pretrain)
+- [Chinese-Roberta-WWM-Ext-Large](https://huggingface.co/hfl/chinese-roberta-wwm-ext-large)
+- [BigVGAN](https://github.com/NVIDIA/BigVGAN)
+
+### 推理用文本前端
+
+- [paddlespeech zh_normalization](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization)
+- [split-lang](https://github.com/DoodleBears/split-lang)
+- [g2pW](https://github.com/GitYCC/g2pW)
+- [pypinyin-g2pW](https://github.com/mozillazg/pypinyin-g2pW)
+- [paddlespeech g2pw](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/g2pw)
+
+### WebUI 工具
+
+- [ultimatevocalremovergui](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [SubFix](https://github.com/cronrpc/SubFix)
+- [FFmpeg](https://github.com/FFmpeg/FFmpeg)
+- [gradio](https://github.com/gradio-app/gradio)
+- [faster-whisper](https://github.com/SYSTRAN/faster-whisper)
+- [FunASR](https://github.com/alibaba-damo-academy/FunASR)
+- [AP-BWE](https://github.com/yxlu-0102/AP-BWE)
+
+感谢 @Naozumi520 提供粤语训练集, 并在粤语相关知识方面给予指导.
+
+## 感谢所有贡献者的努力
+
+<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
+  <img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
+</a>
diff --git a/docs/en/Changelog_EN.md b/docs/en/Changelog_EN.md
new file mode 100644
index 0000000000000000000000000000000000000000..3c3fe18a975c3f552d83fc6f5d79843ad92aff66
--- /dev/null
+++ b/docs/en/Changelog_EN.md
@@ -0,0 +1,222 @@
+### 20240121 Update
+
+1. Added `is_share` to the `config`. In scenarios like Colab, this can be set to `True` to map the WebUI to the public network.
+2. Added English system translation support to WebUI.
+3. The `cmd-asr` automatically detects if the FunASR model is included; if not found in the default directory, it will be downloaded from ModelScope.
+4. Attempted to fix the SoVITS training ZeroDivisionError reported in [Issue 79](https://github.com/RVC-Boss/GPT-SoVITS/issues/79) by filtering samples with zero length, etc.
+5. Cleaned up cached audio files and other files in the `TEMP` folder.
+6. Significantly reduced the issue of synthesized audio containing the end of the reference audio.
+
+### 20240122 Update
+
+1. Fixed the issue where excessively short output files resulted in repeating the reference audio.
+2. Tested native support for English and Japanese training (Japanese training requires the root directory to be free of non-English special characters).
+3. Improved audio path checking. If an attempt is made to read from an incorrect input path, it will report that the path does not exist instead of an ffmpeg error.
+
+### 20240123 Update
+
+1. Resolved the issue where Hubert extraction caused NaN errors, leading to SoVITS/GPT training ZeroDivisionError.
+2. Added support for quick model switching in the inference WebUI.
+3. Optimized the model file sorting logic.
+4. Replaced `jieba` with `jieba_fast` for Chinese word segmentation.
+
+### 20240126 Update
+
+1. Added support for Chinese-English mixed and Japanese-English mixed output texts.
+2. Added an optional segmentation mode for output.
+3. Fixed the issue of UVR5 reading and automatically jumping out of directories.
+4. Fixed multiple newline issues causing inference errors.
+5. Removed redundant logs in the inference WebUI.
+6. Supported training and inference on Mac.
+7. Automatically forced single precision for GPU that do not support half precision; enforced single precision under CPU inference.
+
+### 20240128 Update
+
+1. Fixed the issue with the pronunciation of numbers converting to Chinese characters.
+2. Fixed the issue of swallowing a few characters at the beginning of sentences.
+3. Excluded unreasonable reference audio lengths by setting restrictions.
+4. Fixed the issue where GPT training did not save checkpoints.
+5. Completed model downloading process in the Dockerfile.
+
+### 20240129 Update
+
+1. Changed training configurations to single precision for GPUs like the 16 series, which have issues with half precision training.
+2. Tested and updated the available Colab version.
+3. Fixed the issue of git cloning the ModelScope FunASR repository with older versions of FunASR causing interface misalignment errors.
+
+### 20240130 Update
+
+1. Automatically removed double quotes from all path-related entries to prevent errors from novice users copying paths with double quotes.
+2. Fixed issues with splitting Chinese and English punctuation and added punctuation at the beginning and end of sentences.
+3. Added splitting by punctuation.
+
+### 20240201 Update
+
+1. Fixed the UVR5 format reading error causing separation failures.
+2. Supported automatic segmentation and language recognition for mixed Chinese-Japanese-English texts.
+
+### 20240202 Update
+
+1. Fixed the issue where an ASR path ending with `/` caused an error in saving the filename.
+2. [PR 377](https://github.com/RVC-Boss/GPT-SoVITS/pull/377) introduced PaddleSpeech's Normalizer to fix issues like reading "xx.xx%" (percent symbols) and "元/吨" being read as "元吨" instead of "元每吨", and fixed underscore errors.
+
+### 20240207 Update
+
+1. Corrected language parameter confusion causing decreased Chinese inference quality reported in [Issue 391](https://github.com/RVC-Boss/GPT-SoVITS/issues/391).
+2. [PR 403](https://github.com/RVC-Boss/GPT-SoVITS/pull/403) adapted UVR5 to higher versions of librosa.
+3. [Commit 14a2851](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8) fixed UVR5 inf everywhere error caused by `is_half` parameter not converting to boolean, resulting in constant half precision inference, which caused `inf` on 16 series GPUs.
+4. Optimized English text frontend.
+5. Fixed Gradio dependencies.
+6. Supported automatic reading of `.list` full paths if the root directory is left blank during dataset preparation.
+7. Integrated Faster Whisper ASR for Japanese and English.
+
+### 20240208 Update
+
+1. [Commit 59f35ad](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b) attempted to fix GPT training hang on Windows 10 1909 and [Issue 232](https://github.com/RVC-Boss/GPT-SoVITS/issues/232) (Traditional Chinese System Language).
+
+### 20240212 Update
+
+1. Optimized logic for Faster Whisper and FunASR, switching Faster Whisper to mirror downloads to avoid issues with Hugging Face connections.
+2. [PR 457](https://github.com/RVC-Boss/GPT-SoVITS/pull/457) enabled experimental DPO Loss training option to mitigate GPT repetition and missing characters by constructing negative samples during training and made several inference parameters available in the inference WebUI.
+
+### 20240214 Update
+
+1. Supported Chinese experiment names in training (previously caused errors).
+2. Made DPO training an optional feature instead of mandatory. If selected, the batch size is automatically halved. Fixed issues with new parameters not being passed in the inference WebUI.
+
+### 20240216 Update
+
+1. Supported input without reference text.
+2. Fixed bugs in Chinese frontend reported in [Issue 475](https://github.com/RVC-Boss/GPT-SoVITS/issues/475).
+
+### 20240221 Update
+
+1. Added a noise reduction option during data processing (noise reduction leaves only 16kHz sampling rate; use only if the background noise is significant).
+2. [PR 559](https://github.com/RVC-Boss/GPT-SoVITS/pull/559), [PR 556](https://github.com/RVC-Boss/GPT-SoVITS/pull/556), [PR 532](https://github.com/RVC-Boss/GPT-SoVITS/pull/532), [PR 507](https://github.com/RVC-Boss/GPT-SoVITS/pull/507), [PR 509](https://github.com/RVC-Boss/GPT-SoVITS/pull/509) optimized Chinese and Japanese frontend processing.
+3. Switched Mac CPU inference to use CPU instead of MPS for faster performance.
+4. Fixed Colab public URL issue.
+
+### 20240306 Update
+
+1. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) accelerated inference by 50% (tested on RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39) .
+2. No longer requires downloading the Chinese FunASR model first when using Faster Whisper non-Chinese ASR.
+3. [PR 610](https://github.com/RVC-Boss/GPT-SoVITS/pull/610) fixed UVR5 reverb removal model where the setting was reversed.
+4. [PR 675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675) enabled automatic CPU inference for Faster Whisper if no CUDA is available.
+5. [PR 573](https://github.com/RVC-Boss/GPT-SoVITS/pull/573) modified `is_half` check to ensure proper CPU inference on Mac.
+
+### 202403/202404/202405 Update
+
+#### Minor Fixes:
+
+1. Fixed issues with the no-reference text mode.
+2. Optimized the Chinese and English text frontend.
+3. Improved API format.
+4. Fixed CMD format issues.
+5. Added error prompts for unsupported languages during training data processing.
+6. Fixed the bug in Hubert extraction.
+
+#### Major Fixes:
+
+1. Fixed the issue of SoVITS training without freezing VQ (which could cause quality degradation).
+2. Added a quick inference branch.
+
+### 20240610 Update
+
+#### Minor Fixes:
+
+1. [PR 1168](https://github.com/RVC-Boss/GPT-SoVITS/pull/1168) & [PR 1169](https://github.com/RVC-Boss/GPT-SoVITS/pull/1169) improved the logic for pure punctuation and multi-punctuation text input.
+2. [Commit 501a74a](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232) fixed CMD format for MDXNet de-reverb in UVR5, supporting paths with spaces.
+3. [PR 1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159) fixed progress bar logic for SoVITS training in `s2_train.py`.
+
+#### Major Fixes:
+
+4. [Commit 99f09c8](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a) fixed the issue of WebUI's GPT fine-tuning not reading BERT feature of Chinese input texts, causing inconsistency with inference and potential quality degradation.
+   **Caution: If you have previously fine-tuned with a large amount of data, it is recommended to retune the model to improve quality.**
+
+### 20240706 Update
+
+#### Minor Fixes:
+
+1. [Commit 1250670](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041) fixed default batch size decimal issue in CPU inference.
+2. [PR 1258](https://github.com/RVC-Boss/GPT-SoVITS/pull/1258), [PR 1265](https://github.com/RVC-Boss/GPT-SoVITS/pull/1265), [PR 1267](https://github.com/RVC-Boss/GPT-SoVITS/pull/1267) fixed issues where denoising or ASR encountering exceptions would exit all pending audio files.
+3. [PR 1253](https://github.com/RVC-Boss/GPT-SoVITS/pull/1253) fixed the issue of splitting decimals when splitting by punctuation.
+4. [Commit a208698](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca) fixed multi-process save logic for multi-GPU training.
+5. [PR 1251](https://github.com/RVC-Boss/GPT-SoVITS/pull/1251) removed redundant `my_utils`.
+
+#### Major Fixes:
+
+6. The accelerated inference code from [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) has been validated and merged into the main branch, ensuring consistent inference effects with the base.
+   It also supports accelerated inference in no-reference text mode.
+
+**Future updates will continue to verify the consistency of changes in the `fast_inference` branch**.
+
+### 20240727 Update
+
+#### Minor Fixes:
+
+1. [PR 1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298) cleaned up redundant i18n code.
+2. [PR 1299](https://github.com/RVC-Boss/GPT-SoVITS/pull/1299) fixed issues where trailing slashes in user file paths caused command line errors.
+3. [PR 756](https://github.com/RVC-Boss/GPT-SoVITS/pull/756) fixed the step calculation logic in GPT training.
+
+#### Major Fixes:
+
+4. [Commit 9588a3c](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2) supported speech rate adjustment for synthesis.
+   Enabled freezing randomness while only adjusting the speech rate.
+
+### 20240806 Update
+
+1. [PR 1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR 1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) Added support for the BS RoFormer vocal accompaniment separation model. [Commit e62e965](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c) Enabled FP16 inference.
+2. Improved Chinese text frontend.
+   - [PR 488](https://github.com/RVC-Boss/GPT-SoVITS/pull/488) added support for polyphonic characters (v2 only);
+   - [PR 987](https://github.com/RVC-Boss/GPT-SoVITS/pull/987) added quantifier;
+   - [PR 1351](https://github.com/RVC-Boss/GPT-SoVITS/pull/1351) supports arithmetic and basic math formulas;
+   - [PR 1404](https://github.com/RVC-Boss/GPT-SoVITS/pull/1404) fixed mixed text errors.
+3. [PR 1355](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) automatically filled in the paths when processing audio in the WebUI.
+4. [Commit bce451a](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299), [Commit 4c8b761](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78) optimized GPU recognition logic.
+5. [Commit 8a10147](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3) added support for Cantonese ASR.
+6. Added support for GPT-SoVITS v2.
+7. [PR 1387](https://github.com/RVC-Boss/GPT-SoVITS/pull/1387) optimized timing logic.
+
+### 20240821 Update
+
+1. [PR 1490](https://github.com/RVC-Boss/GPT-SoVITS/pull/1490) Merge the `fast_inference` branch into the main branch.
+2. [Issue 1508](https://github.com/RVC-Boss/GPT-SoVITS/issues/1508) Support for optimizing numbers, phone numbers, dates, and times using SSML tags.
+3. [PR 1503](https://github.com/RVC-Boss/GPT-SoVITS/pull/1503) Fixed and optimized API.
+4. [PR 1422](https://github.com/RVC-Boss/GPT-SoVITS/pull/1422) Fixed the bug where only one reference audio could be uploaded for mixing, Added various dataset checks with warnings popping up if missing files.
+
+### 20250211 Update
+
+- [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)) Added GPT-SoVITS v3 Model, Need 14GB GPU Memory to Fine-tune SoVITS v3.
+
+### 20250212 Update
+
+- [PR 2040](https://github.com/RVC-Boss/GPT-SoVITS/pull/2040) Added gradient checkpointing to Fine-tune SoVITS v3, Need 12GB GPU Memory.
+
+### 20250214 Update
+
+- [PR 2047](https://github.com/RVC-Boss/GPT-SoVITS/pull/2047) Optimize the multilingual mixed text segmentation strategy **A**.
+  -AAdded `split-lang` as a language segmentation tool to improve segmentation capabilities for multi-language mixed text.
+
+### 20250217 Update
+
+- [PR 2062](https://github.com/RVC-Boss/GPT-SoVITS/pull/2062) Optimize the logic for handling numbers and English in the text.
+
+### 20250218 Update
+
+- [PR 2073](https://github.com/RVC-Boss/GPT-SoVITS/pull/2073) Optimize the multilingual mixed text segmentation strategy **B**.
+
+### 20250223 Update
+
+1. LoRA training is supported for fine-tuning with SoVITS V3. It requires 8GB GPU Memory and the results are better than full parameter fine-tuning.
+2. [PR 2078](https://github.com/RVC-Boss/GPT-SoVITS/pull/2078) Added Mel Band RoFormer model for Vocal & Instrument Separation.
+
+### 20250226 Update
+
+1. [PR 2112](https://github.com/RVC-Boss/GPT-SoVITS/pull/2112) Fix issues caused by non-English directories in Windows.
+   - Using `langsegmenter` for Korean.
+2. [PR 2113](https://github.com/RVC-Boss/GPT-SoVITS/pull/2114) Fix issues caused by non-English directories in Windows.
+   - Using `langsegmenter` for Korean/Japanese.
+
+### 20250227 Update
+
+- Added 24K to 48K audio super-resolution models to alleviate the muffled issue when generating 24K audio with V3 model, as reported in [Issue 2085](https://github.com/RVC-Boss/GPT-SoVITS/issues/2085), [Issue 2117](https://github.com/RVC-Boss/GPT-SoVITS/issues/2117).
\ No newline at end of file
diff --git a/docs/ja/Changelog_JA.md b/docs/ja/Changelog_JA.md
new file mode 100644
index 0000000000000000000000000000000000000000..1173c081583fd520b9ef317eeec406c0e9033c35
--- /dev/null
+++ b/docs/ja/Changelog_JA.md
@@ -0,0 +1,221 @@
+### 20240121 更新
+
+1. `config`に`is_share`を追加し、Colab などの環境でこれを`True`に設定すると、webui を公共ネットワークにマッピングできます.
+2. WebUI に英語システムの英語翻訳を追加しました.
+3. `cmd-asr`は FunASR モデルが既に含まれているかどうかを自動的に確認し、デフォルトのパスにない場合は modelscope から自動的にダウンロードします.
+4. [SoVITS 训练报错 ZeroDivisionError](https://github.com/RVC-Boss/GPT-SoVITS/issues/79) 修復を試みます (長さ 0 のサンプルをフィルタリングなど)
+5. TEMP ファイルフォルダからオーディオやその他のファイルをクリーンアップして最適化します.
+6. 合成オーディオがリファレンスオーディオの終わりを含む問題を大幅に改善しました.
+
+### 20240122 更新
+
+1. 短すぎる出力ファイルが重複したリファレンスオーディオを返す問題を修正しました.
+2. 英語-日本語学習がスムーズに進む QA を完了しました. (ただし、日本語学習はルートディレクトリに英語以外の文字が含まれていない必要があります)
+3. オーディオパスをチェックします.間違ったパスを読み取ろうとすると、「パスが存在しません」というエラーメッセージが返されます.これは ffmpeg モジュールのエラーではありません.
+
+### 20240123 更新
+
+1. hubert から nan 抽出による SoVITS/GPT 学習中の ZeroDivisionError 関連エラーを修正しました.
+2. 推論インターフェースでモデルを素早く切り替えることができるようにサポートしました.
+3. モデルファイルのソートロジックを最適化しました.
+4. 中国語の分析に `jieba_fast` を `jieba` に置き換えました.
+
+### 20240126 更新
+
+1. 中国語と英語、日本語と英語が混在した出力テキストをサポートします.
+2. 出力で選択的な分割モードをサポートします.
+3. uvr5 がディレクトリを読み取り、自動的に終了する問題を修正しました.
+4. 複数の改行による推論エラーを修正しました.
+5. 推論インターフェースから不要なログを削除しました.
+6. MacOS での学習と推論をサポートします.
+7. 半精度をサポートしていないカードを自動的に識別して単精度を強制し、CPU 推論では単精度を強制します.
+
+### 20240128 更新
+
+1. 数字を漢字で読む問題を修正しました.
+2. 文章の先頭の一部の単語が欠落する問題を修正しました.
+3. 不適切な長さのリファレンスオーディオを制限しました.
+4. GPT 学習時の ckpt が保存されない問題を修正しました.
+5. Dockerfile のモデルダウンロードプロセスを改善しました.
+
+### 20240129 更新
+
+1. 16 系などの半精度学習に問題があるカードは、学習構成を単精度学習に変更しました.
+2. Colab でも使用可能なバージョンをテストして更新しました.
+3. ModelScope FunASR リポジトリの古いバージョンで git クローンを行う際のインターフェース不整合エラーの問題を修正しました.
+
+### 20240130 更新
+
+1. パスと関連する文字列を解析して、二重引用符を自動的に削除します.また、パスをコピーする場合、二重引用符が含まれていてもエラーが発生しません.
+2. 中国語と英語、日本語と英語の混合出力をサポートします.
+3. 出力で選択的な分割モードをサポートします.
+
+### 20240201 更新
+
+1. UVR5 形式の読み取りエラーによる分離失敗を修正しました.
+2. 中国語・日本語・英語の混合テキストに対する自動分割と言語認識をサポートしました.
+
+### 20240202 更新
+
+1. ASRパスが `/` で終わることによるファイル名保存エラーの問題を修正しました.
+2. [PR 377](https://github.com/RVC-Boss/GPT-SoVITS/pull/377) で PaddleSpeech の Normalizer を導入し、"xx.xx%" (パーセント記号) の読み取りや"元/吨"が"元吨"ではなく"元每吨"と読まれる問題、アンダースコアエラーを修正しました.
+
+### 20240207 更新
+
+1. [Issue 391](https://github.com/RVC-Boss/GPT-SoVITS/issues/391) で報告された中国語推論品質の低下を引き起こした言語パラメータの混乱を修正しました.
+2. [PR 403](https://github.com/RVC-Boss/GPT-SoVITS/pull/403) で UVR5 を librosa のより高いバージョンに適応させました.
+3. [Commit 14a2851](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8) で、`is_half` パラメータがブール値に変換されず、常に半精度推論が行われ、16 シリーズの GPU で `inf` が発生する UVR5 inf everywhereエラーを修正しました.
+4. 英語テキストフロントエンドを最適化しました.
+5. Gradio の依存関係を修正しました.
+6. データセット準備中にルートディレクトリが空白の場合、`.list` フルパスの自動読み取りをサポートしました.
+7. 日本語と英語のために Faster Whisper ASR を統合しました.
+
+### 20240208 更新
+
+1. [Commit 59f35ad](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b) で、Windows 10 1909 および [Issue 232](https://github.com/RVC-Boss/GPT-SoVITS/issues/232) (繁体字中国語システム言語) での GPT トレーニングのハングを修正する試みを行いました.
+
+### 20240212 更新
+
+1. Faster Whisper と FunASR のロジックを最適化し、Faster Whisper をミラーダウンロードに切り替えて Hugging Face の接続問題を回避しました.
+2. [PR 457](https://github.com/RVC-Boss/GPT-SoVITS/pull/457) で、GPT の繰り返しと文字欠落を軽減するために、トレーニング中に負のサンプルを構築する実験的なDPO Lossトレーニングオプションを有効にし、いくつかの推論パラメータを推論WebUIで利用可能にしました.
+
+### 20240214 更新
+
+1. トレーニングで中国語の実験名をサポート (以前はエラーが発生していました).
+2. DPOトレーニングを必須ではなくオプション機能に変更.選択された場合、バッチサイズは自動的に半分になります.推論 WebUI で新しいパラメータが渡されない問題を修正しました.
+
+### 20240216 更新
+
+1. 参照テキストなしでの入力をサポート.
+2. [Issue 475](https://github.com/RVC-Boss/GPT-SoVITS/issues/475) で報告された中国語フロントエンドのバグを修正しました.
+
+### 20240221 更新
+
+1. データ処理中のノイズ低減オプションを追加 (ノイズ低減は16kHzサンプリングレートのみを残します；背景ノイズが大きい場合にのみ使用してください).
+2. [PR 559](https://github.com/RVC-Boss/GPT-SoVITS/pull/559), [PR 556](https://github.com/RVC-Boss/GPT-SoVITS/pull/556), [PR 532](https://github.com/RVC-Boss/GPT-SoVITS/pull/532), [PR 507](https://github.com/RVC-Boss/GPT-SoVITS/pull/507), [PR 509](https://github.com/RVC-Boss/GPT-SoVITS/pull/509) で中国語と日本語のフロントエンド処理を最適化しました.
+3. Mac CPU 推論を MPS ではなく CPU を使用するように切り替え、パフォーマンスを向上させました.
+4. Colab のパブリック URL の問題を修正しました.
+### 20240306 更新
+
+1. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) で推論速度を50%向上させました (RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39 でテスト).
+2. Faster Whisper非中国語ASRを使用する際、最初に中国語FunASRモデルをダウンロードする必要がなくなりました.
+3. [PR 610](https://github.com/RVC-Boss/GPT-SoVITS/pull/610) で UVR5 残響除去モデルの設定が逆になっていた問題を修正しました.
+4. [PR 675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675) で、CUDA が利用できない場合に Faster Whisper の自動 CPU 推論を有効にしました.
+5. [PR 573](https://github.com/RVC-Boss/GPT-SoVITS/pull/573) で、Mac での適切なCPU推論を確保するために `is_half` チェックを修正しました.
+
+### 202403/202404/202405 更新
+
+#### マイナー修正:
+
+1. 参照テキストなしモードの問題を修正しました.
+2. 中国語と英語のテキストフロントエンドを最適化しました.
+3. API フォーマットを改善しました.
+4. CMD フォーマットの問題を修正しました.
+5. トレーニングデータ処理中のサポートされていない言語に対するエラープロンプトを追加しました.
+6. Hubert 抽出のバグを修正しました.
+
+#### メジャー修正:
+
+1. SoVITS トレーニングで VQ を凍結せずに品質低下を引き起こす問題を修正しました.
+2. クイック推論ブランチを追加しました.
+
+### 20240610 更新
+
+#### マイナー修正:
+
+1. [PR 1168](https://github.com/RVC-Boss/GPT-SoVITS/pull/1168) & [PR 1169](https://github.com/RVC-Boss/GPT-SoVITS/pull/1169)で、純粋な句読点および複数の句読点を含むテキスト入力のロジックを改善しました.
+2. [Commit 501a74a](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232)で、UVR5 の MDXNet デリバブをサポートする CMD フォーマットを修正し、スペースを含むパスをサポートしました.
+3. [PR 1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159)で、`s2_train.py` の SoVITS トレーニングのプログレスバーロジックを修正しました.
+
+#### メジャー修正:
+
+4. [Commit 99f09c8](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a) で、WebUI の GPT ファインチューニングが中国語入力テキストの BERT 特徴を読み取らず、推論との不一致や品質低下の可能性を修正しました.
+   **注意: 以前に大量のデータでファインチューニングを行った場合、品質向上のためにモデルを再調整することをお勧めします.**
+
+### 20240706 更新
+
+#### マイナー修正:
+
+1. [Commit 1250670](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041) で、CPU 推論のデフォルトバッチサイズの小数点問題を修正しました.
+2. [PR 1258](https://github.com/RVC-Boss/GPT-SoVITS/pull/1258), [PR 1265](https://github.com/RVC-Boss/GPT-SoVITS/pull/1265), [PR 1267](https://github.com/RVC-Boss/GPT-SoVITS/pull/1267) で、ノイズ除去またはASRが例外に遭遇した場合に、すべての保留中のオーディオファイルが終了する問題を修正しました.
+3. [PR 1253](https://github.com/RVC-Boss/GPT-SoVITS/pull/1253) で、句読点で分割する際の小数点分割の問題を修正しました.
+4. [Commit a208698](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca) で、マルチGPUトレーニングのマルチプロセス保存ロジックを修正しました.
+5. [PR 1251](https://github.com/RVC-Boss/GPT-SoVITS/pull/1251) で、不要な `my_utils` を削除しました.
+
+#### メジャー修正:
+
+6. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) の加速推論コードが検証され、メインブランチにマージされ、ベースとの推論効果の一貫性が確保されました.
+   また、参照テキストなしモードでの加速推論もサポートしています.
+
+**今後の更新では、`fast_inference`ブランチの変更の一貫性を継続的に検証します**.
+
+### 20240727 更新
+
+#### マイナー修正:
+
+1. [PR 1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298) で、不要な i18n コードをクリーンアップしました.
+2. [PR 1299](https://github.com/RVC-Boss/GPT-SoVITS/pull/1299) で、ユーザーファイルパスの末尾のスラッシュがコマンドラインエラーを引き起こす問題を修正しました.
+3. [PR 756](https://github.com/RVC-Boss/GPT-SoVITS/pull/756) で、GPT トレーニングのステップ計算ロジックを修正しました.
+
+#### メジャー修正:
+
+4. [Commit 9588a3c](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2) で、合成のスピーチレート調整をサポートしました.
+   スピーチレートのみを調整しながらランダム性を固定できるようになりました.
+
+### 20240806 更新
+
+1. [PR 1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306)、[PR 1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) BS RoFormer ボーカルアコムパニ分離モデルのサポートを追加しました.[Commit e62e965](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c) FP16 推論を有効にしました.
+2. 中国語テキストフロントエンドを改善しました.
+   - [PR 488](https://github.com/RVC-Boss/GPT-SoVITS/pull/488) 多音字のサポートを追加 (v2 のみ);
+   - [PR 987](https://github.com/RVC-Boss/GPT-SoVITS/pull/987) 量詞を追加;
+   - [PR 1351](https://github.com/RVC-Boss/GPT-SoVITS/pull/1351) 四則演算と基本数式のサポート;
+   - [PR 1404](https://github.com/RVC-Boss/GPT-SoVITS/pull/1404) 混合テキストエラーを修正.
+3. [PR 1355](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) WebUIでオーディオ処理時にパスを自動入力しました.
+4. [Commit bce451a](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299), [Commit 4c8b761](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78) GPU 認識ロジックを最適化しました.
+5. [Commit 8a10147](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3) 広東語ASRのサポートを追加しました.
+6. GPT-SoVITS v2 のサポートを追加しました.
+7. [PR 1387](https://github.com/RVC-Boss/GPT-SoVITS/pull/1387) タイミングロジックを最適化しました.
+
+### 20240821 更新
+
+1. [PR 1490](https://github.com/RVC-Boss/GPT-SoVITS/pull/1490) `fast_inference` ブランチをメインブランチにマージしました.
+2. [Issue 1508](https://github.com/RVC-Boss/GPT-SoVITS/issues/1508) SSMLタグを使用して数字、電話番号、日付、時間などの最適化をサポートしました.
+3. [PR 1503](https://github.com/RVC-Boss/GPT-SoVITS/pull/1503) APIの修正と最適化を行いました.
+4. [PR 1422](https://github.com/RVC-Boss/GPT-SoVITS/pull/1422) 参照音声のミキシングで1つしかアップロードできないバグを修正し、データセットの各種チェックを追加してファイルが欠落している場合に警告を表示するようにしました.
+
+### 20250211 更新
+
+1. [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)) GPT-SoVITS v3 モデルを追加しました.SoVITS v3のファインチューニングには14GBのGPUメモリが必要です.
+
+### 20250212 更新
+
+- [PR 2040](https://github.com/RVC-Boss/GPT-SoVITS/pull/2040) SoVITS v3のファインチューニングにグラデーションチェックポイントを追加、12GBのGPUメモリが必要です.
+
+### 20250214 更新
+
+- [PR 2047](https://github.com/RVC-Boss/GPT-SoVITS/pull/2047) 多言語混合テキスト分割戦略の最適化 **A**.
+  - `split-lang`を言語分割ツールとして追加し、多言語混合テキストの分割能力を向上させました.
+
+### 20250217 更新
+
+- [PR 2062](https://github.com/RVC-Boss/GPT-SoVITS/pull/2062) テキスト内の数字と英語の処理ロジックを最適化.
+
+### 20250218 更新
+
+- [PR 2073](https://github.com/RVC-Boss/GPT-SoVITS/pull/2073) 多言語混合テキスト分割戦略の最適化 **B**.
+
+### 20250223 更新
+
+1. LoRAトレーニングがSoVITS V3のファインチューニングに対応しました.8GBのGPUメモリが必要で、結果はフルパラメータファインチューニングより優れています.
+2. [PR 2078](https://github.com/RVC-Boss/GPT-SoVITS/pull/2078) ボーカルと楽器分離のためにMel Band RoFormerモデルを追加しました.
+
+### 20250226 更新
+
+1. [PR 2112](https://github.com/RVC-Boss/GPT-SoVITS/pull/2112) Windowsでの非英語ディレクトリによる問題を修正しました.
+   - `langsegmenter`を使用して韓国語の問題を修正.
+2. [PR 2113](https://github.com/RVC-Boss/GPT-SoVITS/pull/2114) Windowsでの非英語ディレクトリによる問題を修正しました.
+   - `langsegmenter`を使用して韓国語/日本語の問題を修正.
+
+### 20250227 更新
+
+- V3モデルで24Kオーディオを生成する際に発生するこもった音の問題を緩和するために、24Kから48Kのオーディオ超解像モデルを追加しました.[Issue 2085](https://github.com/RVC-Boss/GPT-SoVITS/issues/2085)、[Issue 2117](https://github.com/RVC-Boss/GPT-SoVITS/issues/2117)で報告されています.
\ No newline at end of file
diff --git a/docs/ja/README.md b/docs/ja/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..145de219dfebd426014883f384dbf053326ab72c
--- /dev/null
+++ b/docs/ja/README.md
@@ -0,0 +1,383 @@
+<div align="center">
+
+<h1>GPT-SoVITS-WebUI</h1>
+パワフルなFew-Shot音声変換・音声合成 WebUI.<br><br>
+
+[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
+
+<img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br>
+
+[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
+[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
+[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
+[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
+
+[**English**](../../README.md) | [**中文简体**](../cn/README.md) | **日本語** | [**한국어**](../ko/README.md) | [**Türkçe**](../tr/README.md)
+
+</div>
+
+---
+
+## 機能:
+
+1. **Zero-Shot TTS:** たった 5 秒間の音声サンプルで、即座にテキストからその音声に変換できます.
+
+2. **Few-Shot TTS:** わずか 1 分間のトレーニングデータでモデルを微調整し、音声のクオリティを向上.
+
+3. **多言語サポート:** 現在、英語、日本語、韓国語、広東語、中国語をサポートしています.
+
+4. **WebUI ツール:** 統合されたツールは、音声と伴奏 (BGM 等) の分離、トレーニングセットの自動セグメンテーション、ASR (中国語のみ)、テキストラベリング等を含むため、初心者の方でもトレーニングデータセットの作成や GPT/SoVITS モデルのトレーニング等を非常に簡単に行えます.
+
+**[デモ動画](https://www.bilibili.com/video/BV12g4y1m7Uw)をチェック！**
+
+声の事前学習無しかつ Few-Shot でトレーニングされたモデルのデモ:
+
+https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
+
+**ユーザーマニュアル: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
+
+## インストール
+
+### テスト済みの環境
+
+| Python Version | PyTorch Version  | Device          |
+|----------------|------------------|-----------------|
+| Python 3.9     | PyTorch 2.0.1    | CUDA 11.8       |
+| Python 3.10.13 | PyTorch 2.1.2    | CUDA 12.3       |
+| Python 3.10.17 | PyTorch 2.5.1    | CUDA 12.4       |
+| Python 3.9     | PyTorch 2.5.1    | Apple silicon   |
+| Python 3.11    | PyTorch 2.6.0    | Apple silicon   |
+| Python 3.9     | PyTorch 2.2.2    | CPU             |
+
+### Windows
+
+Windows ユーザー: (Windows 10 以降でテスト済み)、[統合パッケージをダウンロード](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-v3lora-20250228.7z?download=true)し、解凍後に _go-webui.bat_ をダブルクリックすると、GPT-SoVITS-WebUI が起動します.
+
+### Linux
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### macOS
+
+**注: Mac で GPU を使用して訓練されたモデルは、他のデバイスで訓練されたモデルと比較して著しく品質が低下するため、当面は CPU を使用して訓練することを強く推奨します.**
+
+1. `xcode-select --install` を実行して、Xcode コマンドラインツールをインストールします.
+2. 以下のコマンドを実行してこのプロジェクトをインストールします.
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### 手動インストール
+
+#### FFmpeg をインストールします.
+
+##### Conda ユーザー
+
+```bash
+conda install ffmpeg
+```
+
+##### Ubuntu/Debian ユーザー
+
+```bash
+sudo apt install ffmpeg
+sudo apt install libsox-dev
+conda install -c conda-forge 'ffmpeg<7'
+```
+
+##### Windows ユーザー
+
+[ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) と [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) をダウンロードし、GPT-SoVITS のルートフォルダに置きます.
+
+##### MacOS ユーザー
+
+```bash
+brew install ffmpeg
+```
+
+#### 依存関係をインストールします
+
+```bash
+pip install -r extra-req.txt --no-deps
+pip install -r requirementx.txt
+```
+
+### Docker の使用
+
+#### docker-compose.yaml の設定
+
+0. イメージのタグについて: コードベースの更新が速い割に、イメージのパッケージングとテストが遅いため、[Docker Hub](https://hub.docker.com/r/breakstring/gpt-sovits)(古いバージョン) で現在パッケージされている最新のイメージをご覧になり、ご自身の状況に応じて選択するか、またはご自身のニーズに応じて Dockerfile を使用してローカルでビルドしてください.
+1. 環境変数: 
+
+   - `is_half`: 半精度／倍精度の制御."SSL 抽出"ステップ中に`4-cnhubert/5-wav32k`ディレクトリ内の内容が正しく生成されない場合、通常これが原因です.実際の状況に応じて True または False に調整してください.
+
+2. ボリューム設定: コンテナ内のアプリケーションのルートディレクトリは`/workspace`に設定されます.デフォルトの`docker-compose.yaml`には、アップロード／ダウンロードの内容の実例がいくつか記載されています.
+3. `shm_size`: Windows の Docker Desktop のデフォルトの利用可能メモリは小さすぎるため、うまく動作しない可能性があります.状況に応じて適宜設定してください.
+4. `deploy`セクションの GPU に関連する内容は、システムと実際の状況に応じて慎重に設定してください.
+
+#### docker compose で実行する
+
+```markdown
+docker compose -f "docker-compose.yaml" up -d
+```
+
+#### docker コマンドで実行する
+
+上記と同様に、実際の状況に基づいて対応するパラメータを変更し、次のコマンドを実行します: 
+
+```markdown
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+```
+
+## 事前訓練済みモデル
+
+**`install.sh`が正常に実行された場合、No.1,2,3 はスキップしてかまいません.**
+
+1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) から事前訓練済みモデルをダウンロードし、`GPT_SoVITS/pretrained_models` ディレクトリに配置してください.
+
+2. [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) からモデルをダウンロードし、解凍して `G2PWModel` にリネームし、`GPT_SoVITS/text` ディレクトリに配置してください. (中国語 TTS のみ)
+
+3. UVR5 (ボーカル/伴奏 (BGM 等) 分離 & リバーブ除去の追加機能) の場合は、[UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) からモデルをダウンロードし、`tools/uvr5/uvr5_weights` ディレクトリに配置してください.
+
+   - UVR5 で bs_roformer または mel_band_roformer モデルを使用する場合、モデルと対応する設定ファイルを手動でダウンロードし、`tools/UVR5/UVR5_weights`フォルダに配置することができます.**モデルファイルと設定ファイルの名前は、拡張子を除いて同じであることを確認してください**.さらに、モデルと設定ファイルの名前には**「roformer」が含まれている必要があります**.これにより、roformer クラスのモデルとして認識されます.
+
+   - モデル名と設定ファイル名には、**直接モデルタイプを指定することをお勧めします**.例: mel_mand_roformer、bs_roformer.指定しない場合、設定文から特徴を照合して、モデルの種類を特定します.例えば、モデル`bs_roformer_ep_368_sdr_12.9628.ckpt`と対応する設定ファイル`bs_roformer_ep_368_sdr_12.9628.yaml`はペアです.同様に、`kim_mel_band_roformer.ckpt`と`kim_mel_band_roformer.yaml`もペアです.
+
+4. 中国語 ASR (追加機能) の場合は、[Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files)、[Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files)、および [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) からモデルをダウンロードし、`tools/asr/models` ディレクトリに配置してください.
+
+5. 英語または日本語の ASR (追加機能) を使用する場合は、[Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) からモデルをダウンロードし、`tools/asr/models` ディレクトリに配置してください.また、[他のモデル](https://huggingface.co/Systran) は、より小さいサイズで高クオリティな可能性があります.
+
+## データセット形式
+
+TTS アノテーション .list ファイル形式:
+
+```
+vocal_path|speaker_name|language|text
+```
+
+言語辞書:
+
+- 'zh': 中国語
+- 'ja': 日本語
+- 'en': 英語
+
+例:
+
+```
+D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
+```
+
+## 微調整と推論
+
+### WebUI を開く
+
+#### 統合パッケージ利用者
+
+`go-webui.bat`をダブルクリックするか、`go-webui.ps1`を使用します.
+V1 に切り替えたい場合は、`go-webui-v1.bat`をダブルクリックするか、`go-webui-v1.ps1`を使用してください.
+
+#### その他
+
+```bash
+python webui.py <言語(オプション)>
+```
+
+V1 に切り替えたい場合は
+
+```bash
+python webui.py v1 <言語(オプション)>
+```
+
+または WebUI で手動でバージョンを切り替えてください.
+
+### 微調整
+
+#### パス自動補完のサポート
+
+    1. 音声パスを入力する
+    2. 音声を小さなチャンクに分割する
+    3. ノイズ除去 (オプション)
+    4. ASR
+    5. ASR転写を校正する
+    6. 次のタブに移動し、モデルを微調整する
+
+### 推論 WebUI を開く
+
+#### 統合パッケージ利用者
+
+`go-webui-v2.bat`をダブルクリックするか、`go-webui-v2.ps1`を使用して、`1-GPT-SoVITS-TTS/1C-inference`で推論 webui を開きます.
+
+#### その他
+
+```bash
+python GPT_SoVITS/inference_webui.py <言語(オプション)>
+```
+
+または
+
+```bash
+python webui.py
+```
+
+その後、`1-GPT-SoVITS-TTS/1C-inference`で推論 webui を開きます.
+
+## V2 リリースノート
+
+新機能:
+
+1. 韓国語と広東語をサポート
+
+2. 最適化されたテキストフロントエンド
+
+3. 事前学習済みモデルが 2 千時間から 5 千時間に拡張
+
+4. 低品質の参照音声に対する合成品質の向上
+
+   [詳細はこちら](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v2%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+V1 環境から V2 を使用するには:
+
+1. `pip install -r requirements.txt`を使用していくつかのパッケージを更新
+
+2. 最新のコードを github からクローン
+
+3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained)から V2 の事前学習モデルをダウンロードし、それらを`GPT_SoVITS\pretrained_models\gsv-v2final-pretrained`に配置
+
+   中国語 V2 追加: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW モデルをダウンロードし、解凍して`G2PWModel`にリネームし、`GPT_SoVITS/text`に配置します)
+
+## V3 リリースノート
+
+新機能:
+
+1. 音色の類似性が向上し、ターゲットスピーカーを近似するために必要な学習データが少なくなりました (音色の類似性は、ファインチューニングなしでベースモデルを直接使用することで顕著に改善されます).
+
+2. GPT モデルがより安定し、繰り返しや省略が減少し、より豊かな感情表現を持つ音声の生成が容易になりました.
+
+   [詳細情報はこちら](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+v2 環境から v3 を使用する方法:
+
+1. `pip install -r requirements.txt` を実行して、いくつかのパッケージを更新します.
+
+2. GitHub から最新のコードをクローンします.
+
+3. v3 の事前学習済みモデル (s1v3.ckpt、s2Gv3.pth、models--nvidia--bigvgan_v2_24khz_100band_256x フォルダ) を[Huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) からダウンロードし、GPT_SoVITS\pretrained_models フォルダに配置します.
+
+   追加: 音声超解像モデルについては、[ダウンロード方法](../../tools/AP_BWE_main/24kto48k/readme.txt)を参照してください.
+
+## Todo リスト
+
+- [x] **優先度 高:**
+
+  - [x] 日本語と英語でのローカライズ.
+  - [x] ユーザーガイド.
+  - [x] 日本語データセットと英語データセットのファインチューニングトレーニング.
+
+- [ ] **機能:**
+  - [x] ゼロショット音声変換 (5 秒) ／数ショット音声変換 (1 分).
+  - [x] TTS スピーキングスピードコントロール.
+  - [ ] ~~TTS の感情コントロールの強化.~~
+  - [ ] SoVITS トークン入力を語彙の確率分布に変更する実験.
+  - [x] 英語と日本語のテキストフロントエンドを改善.
+  - [ ] 小型と大型の TTS モデルを開発する.
+  - [x] Colab のスクリプト.
+  - [ ] トレーニングデータセットを拡張する (2k→10k).
+  - [x] より良い sovits ベースモデル (音質向上)
+  - [ ] モデルミックス
+
+## (追加の) コマンドラインから実行する方法
+
+コマンド ラインを使用して UVR5 の WebUI を開きます
+
+```
+python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
+```
+
+<!-- ブラウザを開けない場合は、以下の形式に従って UVR 処理を行ってください.これはオーディオ処理に mdxnet を使用しています.
+```
+python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
+``` -->
+
+コマンド ラインを使用してデータセットのオーディオ セグメンテーションを行う方法は次のとおりです.
+
+```
+python audio_slicer.py \
+    --input_path "<path_to_original_audio_file_or_directory>" \
+    --output_root "<directory_where_subdivided_audio_clips_will_be_saved>" \
+    --threshold <volume_threshold> \
+    --min_length <minimum_duration_of_each_subclip> \
+    --min_interval <shortest_time_gap_between_adjacent_subclips>
+    --hop_size <step_size_for_computing_volume_curve>
+```
+
+コマンドラインを使用してデータセット ASR 処理を行う方法です (中国語のみ)
+
+```
+python tools/asr/funasr_asr.py -i <input> -o <output>
+```
+
+ASR 処理は Faster_Whisper を通じて実行されます(中国語を除く ASR マーキング)
+
+(進行状況バーは表示されません.GPU のパフォーマンスにより時間遅延が発生する可能性があります)
+
+```
+python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p <precision>
+```
+
+カスタムリストの保存パスが有効になっています
+
+## クレジット
+
+特に以下のプロジェクトと貢献者に感謝します: 
+
+### 理論研究
+
+- [ar-vits](https://github.com/innnky/ar-vits)
+- [SoundStorm](https://github.com/yangdongchao/SoundStorm/tree/master/soundstorm/s1/AR)
+- [vits](https://github.com/jaywalnut310/vits)
+- [TransferTTS](https://github.com/hcy71o/TransferTTS/blob/master/models.py#L556)
+- [contentvec](https://github.com/auspicious3000/contentvec/)
+- [hifi-gan](https://github.com/jik876/hifi-gan)
+- [fish-speech](https://github.com/fishaudio/fish-speech/blob/main/tools/llama/generate.py#L41)
+- [f5-TTS](https://github.com/SWivid/F5-TTS/blob/main/src/f5_tts/model/backbones/dit.py)
+- [shortcut flow matching](https://github.com/kvfrans/shortcut-models/blob/main/targets_shortcut.py)
+
+### 事前学習モデル
+
+- [Chinese Speech Pretrain](https://github.com/TencentGameMate/chinese_speech_pretrain)
+- [Chinese-Roberta-WWM-Ext-Large](https://huggingface.co/hfl/chinese-roberta-wwm-ext-large)
+- [BigVGAN](https://github.com/NVIDIA/BigVGAN)
+
+### 推論用テキストフロントエンド
+
+- [paddlespeech zh_normalization](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization)
+- [split-lang](https://github.com/DoodleBears/split-lang)
+- [g2pW](https://github.com/GitYCC/g2pW)
+- [pypinyin-g2pW](https://github.com/mozillazg/pypinyin-g2pW)
+- [paddlespeech g2pw](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/g2pw)
+
+### WebUI ツール
+
+- [ultimatevocalremovergui](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [SubFix](https://github.com/cronrpc/SubFix)
+- [FFmpeg](https://github.com/FFmpeg/FFmpeg)
+- [gradio](https://github.com/gradio-app/gradio)
+- [faster-whisper](https://github.com/SYSTRAN/faster-whisper)
+- [FunASR](https://github.com/alibaba-damo-academy/FunASR)
+- [AP-BWE](https://github.com/yxlu-0102/AP-BWE)
+
+@Naozumi520 さん、広東語のトレーニングセットの提供と、広東語に関する知識のご指導をいただき、感謝申し上げます.
+
+## すべてのコントリビューターに感謝します
+
+<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
+  <img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
+</a>
diff --git a/docs/ko/Changelog_KO.md b/docs/ko/Changelog_KO.md
new file mode 100644
index 0000000000000000000000000000000000000000..e0cfc3b08b1f509bf1be1ee46eab40012bff53e3
--- /dev/null
+++ b/docs/ko/Changelog_KO.md
@@ -0,0 +1,222 @@
+### 20240121 업데이트
+
+1. `config`에 `is_share`를 추가했습니다. Colab과 같은 시나리오에서는 이 값을 `True`로 설정하여 WebUI를 공개 네트워크에 매핑할 수 있습니다.
+2. WebUI에 영어 시스템 번역 지원을 추가했습니다.
+3. `cmd-asr`이 FunASR 모델이 포함되어 있는지 자동으로 감지합니다; 기본 디렉토리에서 찾을 수 없으면 ModelScope에서 다운로드됩니다.
+4. [Issue 79](https://github.com/RVC-Boss/GPT-SoVITS/issues/79)에서 보고된 SoVITS 훈련의 ZeroDivisionError를 필터링 샘플 등으로 해결하려고 시도했습니다.
+5. `TEMP` 폴더의 캐시된 오디오 파일 및 기타 파일을 정리했습니다.
+6. 참조 오디오의 끝이 포함된 합성 오디오 문제를 크게 줄였습니다.
+
+### 20240122 업데이트
+
+1. 지나치게 짧은 출력 파일로 인해 참조 오디오가 반복되는 문제를 수정했습니다.
+2. 영어 및 일본어 훈련의 네이티브 지원을 테스트했습니다 (일본어 훈련 시 루트 디렉토리에 비영어 특수 문자가 없어야 합니다).
+3. 오디오 경로 확인을 개선했습니다. 잘못된 입력 경로에서 읽으려는 시도가 있을 경우, ffmpeg 오류 대신 경로가 존재하지 않는다고 보고합니다.
+
+### 20240123 업데이트
+
+1. Hubert 추출로 인해 NaN 오류가 발생하여 SoVITS/GPT 훈련에서 ZeroDivisionError가 발생하는 문제를 해결했습니다.
+2. 추론 WebUI에서 빠른 모델 전환 지원을 추가했습니다.
+3. 모델 파일 정렬 로직을 최적화했습니다.
+4. 중국어 단어 분할을 위해 `jieba`를 `jieba_fast`로 교체했습니다.
+
+### 20240126 업데이트
+
+1. 중국어-영어 혼합 및 일본어-영어 혼합 출력 텍스트를 지원합니다.
+2. 출력에 대한 선택적 분할 모드를 추가했습니다.
+3. UVR5 읽기 문제 및 디렉토리 자동 탈출 문제를 수정했습니다.
+4. 추론 오류를 일으키는 여러 줄 바꿈 문제를 수정했습니다.
+5. 추론 WebUI 에서 중복 로그를 제거했습니다.
+6. Mac에서 훈련 및 추론을 지원합니다.
+7. 절반 정밀도를 지원하지 않는 GPU에 대해 자동으로 단정밀도를 강제하며, CPU 추론 시 단정밀도를 적용합니다.
+
+### 20240128 업데이트
+
+1. 숫자의 발음이 중국어 문자로 변환되는 문제를 수정했습니다.
+2. 문장 시작 부분에서 몇 개의 문자가 누락되는 문제를 수정했습니다.
+3. 비합리적인 참조 오디오 길이를 설정하여 제외했습니다.
+4. GPT 훈련 시 체크포인트가 저장되지 않는 문제를 수정했습니다.
+5. Dockerfile 에서 모델 다운로드 프로세스를 완료했습니다.
+
+### 20240129 업데이트
+
+1. 절반 정밀도 훈련에 문제가 있는 16 시리즈와 같은 GPU의 훈련 구성을 단정밀도로 변경했습니다.
+2. 사용 가능한 Colab 버전을 테스트하고 업데이트했습니다.
+3. 이전 버전의 FunASR 로 인해 인터페이스 정렬 오류가 발생하는 ModelScope FunASR 저장소의 git 클로닝 문제를 수정했습니다.
+
+### 20240130 업데이트
+
+1. 모든 경로 관련 항목에서 이중 따옴표를 자동으로 제거하여 초보자가 이중 따옴표가 포함된 경로를 복사하는 오류를 방지했습니다.
+2. 중국어 및 영어 문장 부호 분할 문제를 수정하고 문장 시작과 끝에 부호를 추가했습니다.
+3. 부호에 의한 분할을 추가했습니다.
+
+### 20240201 업데이트
+
+1. 분리 실패를 일으킨 UVR5 형식 읽기 오류를 수정했습니다.
+2. 혼합된 중국어-일본어-영어 텍스트에 대한 자동 분할 및 언어 인식을 지원합니다.
+
+### 20240202 업데이트
+
+1. `/` 로 끝나는 ASR 경로가 파일 이름 저장 시 오류를 발생시키는 문제를 수정했습니다.
+2. [PR 377](https://github.com/RVC-Boss/GPT-SoVITS/pull/377) 에서는 PaddleSpeech 의 Normalizer 를 도입하여 "xx.xx%" (백분율 기호)와 "元/吨"이 "元吨"으로 읽히는 문제를 "元每吨"으로 수정하고, 밑줄 오류를 수정했습니다.
+
+### 20240207 업데이트
+
+1. [Issue 391](https://github.com/RVC-Boss/GPT-SoVITS/issues/391) 에서 보고된 중국어 추론 품질 저하를 일으킨 언어 매개변수 혼동을 수정했습니다.
+2. [PR 403](https://github.com/RVC-Boss/GPT-SoVITS/pull/403) 에서는 UVR5 를 높은 버전의 librosa에 맞게 조정했습니다.
+3. [Commit 14a2851](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8)에서는 `is_half` 매개변수가 불리언으로 변환되지 않아 발생한 UVR5 `inf` 오류를 수정했습니다. 이로 인해 16 시리즈 GPU에서 `inf` 가 발생했습니다.
+4. 영어 텍스트 프론트엔드를 최적화했습니다.
+5. Gradio 종속성 문제를 수정했습니다.
+6. 데이터셋 준비 시 루트 디렉토리를 비워두면 `.list` 전체 경로를 자동으로 읽도록 지원합니다.
+7. 일본어와 영어에 대한 Faster Whisper ASR을 통합했습니다.
+
+### 20240208 업데이트
+
+1. [Commit 59f35ad](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b)에서는 Windows 10 1909와 [Issue 232](https://github.com/RVC-Boss/GPT-SoVITS/issues/232) (전통 중국어 시스템 언어)에서 GPT 훈련 멈춤 문제를 수정하려고 했습니다.
+
+### 20240212 업데이트
+
+1. Faster Whisper와 FunASR의 로직을 최적화하고, Faster Whisper를 미러 다운로드로 전환하여 Hugging Face 연결 문제를 피했습니다.
+2. [PR 457](https://github.com/RVC-Boss/GPT-SoVITS/pull/457)은 DPO Loss 실험적 훈련 옵션을 활성화하여 GPT의 반복 및 문자 누락 문제를 완화하고, 훈련 중 부정 샘플을 구성하며 여러 추론 매개변수를 추론 WebUI에서 사용할 수 있게 했습니다.
+
+### 20240214 업데이트
+
+1. 훈련 시 중국어 실험 이름을 지원합니다 (이전에는 오류가 발생했습니다).
+2. DPO 훈련을 필수 기능 대신 선택적 기능으로 변경했습니다. 선택 시, 배치 크기가 자동으로 절반으로 줄어듭니다. 추론 WebUI에서 새로운 매개변수가 전달되지 않는 문제를 수정했습니다.
+
+### 20240216 업데이트
+
+1. 참조 텍스트 없이 입력을 지원합니다.
+2. [Issue 475](https://github.com/RVC-Boss/GPT-SoVITS/issues/475)에서 보고된 중국어 프론트엔드의 버그를 수정했습니다.
+
+### 20240221 업데이트
+
+1. 데이터 처리 중 노이즈 감소 옵션을 추가했습니다 (노이즈 감소는 16kHz 샘플링 비율만 남깁니다; 배경 노이즈가 심한 경우에만 사용하십시오).
+2. [PR 559](https://github.com/RVC-Boss/GPT-SoVITS/pull/559), [PR 556](https://github.com/RVC-Boss/GPT-SoVITS/pull/556), [PR 532](https://github.com/RVC-Boss/GPT-SoVITS/pull/532), [PR 507](https://github.com/RVC-Boss/GPT-SoVITS/pull/507), [PR 509](https://github.com/RVC-Boss/GPT-SoVITS/pull/509) 중국어 및 일본어 프론트엔드 처리를 최적화했습니다.
+3. Mac CPU 추론을 MPS 대신 CPU를 사용하도록 전환하여 성능을 향상시켰습니다.
+4. Colab 공개 URL 문제를 수정했습니다.
+
+### 20240306 업데이트
+
+1. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672)는 추론 속도를 50% 가속화했습니다 (RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39에서 테스트됨).
+2. Faster Whisper의 비중국어 ASR을 사용할 때 중국어 FunASR 모델을 먼저 다운로드할 필요가 없습니다.
+3. [PR 610](https://github.com/RVC-Boss/GPT-SoVITS/pull/610)은 UVR5 리버브 제거 모델에서 설정이 반대로 되어 있는 문제를 수정했습니다.
+4. [PR 675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675)는 CUDA가 없는 경우 Faster Whisper의 자동 CPU 추론을 가능하게 했습니다.
+5. [PR 573](https://github.com/RVC-Boss/GPT-SoVITS/pull/573)은 Mac에서 올바른 CPU 추론을 보장하기 위해 `is_half` 체크를 수정했습니다.
+
+### 202403/202404/202405 업데이트
+
+#### 사소한 수정:
+
+1. 참조 텍스트 없는 모드의 문제를 수정했습니다.
+2. 중국어 및 영어 텍스트 프론트엔드를 최적화했습니다.
+3. API 형식을 개선했습니다.
+4. CMD 형식 문제를 수정했습니다.
+5. 훈련 데이터 처리 중 지원되지 않는 언어에 대한 오류 프롬프트를 추가했습니다.
+6. Hubert 추출의 버그를 수정했습니다.
+
+#### 주요 수정:
+
+1. VQ를 고정하지 않고 SoVITS 훈련의 문제를 수정했습니다(품질 저하를 일으킬 수 있음).
+2. 빠른 추론 분기를 추가했습니다.
+
+### 20240610 업데이트
+
+#### 사소한 수정:
+
+1. [PR 1168](https://github.com/RVC-Boss/GPT-SoVITS/pull/1168) & [PR 1169](https://github.com/RVC-Boss/GPT-SoVITS/pull/1169) 순수 구두점 및 다중 구두점 텍스트 입력 로직을 개선했습니다.
+2. [Commit 501a74a](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232) UVR5에서 MDXNet 디러버브를 위한 CMD 형식을 수정하고 공백이 있는 경로를 지원했습니다.
+3. [PR 1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159) `s2_train.py`에서 SoVITS 훈련을 위한 진행률 표시줄 로직을 수정했습니다.
+
+#### 주요 수정:
+
+4. [Commit 99f09c8](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a) WebUI의 GPT 미세 조정이 중국어 입력 텍스트의 BERT 기능을 읽지 않아 추론과 불일치 및 잠재적 품질 저하를 일으키는 문제를 수정했습니다.
+   **주의: 이전에 많은 양의 데이터로 미세 조정한 경우 품질을 향상시키기 위해 모델을 다시 조정하는 것이 좋습니다.**
+
+### 20240706 업데이트
+
+#### 사소한 수정:
+
+1. [Commit 1250670](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041) CPU 추론에서 기본 배치 크기 소수점 문제를 수정했습니다.
+2. [PR 1258](https://github.com/RVC-Boss/GPT-SoVITS/pull/1258), [PR 1265](https://github.com/RVC-Boss/GPT-SoVITS/pull/1265), [PR 1267](https://github.com/RVC-Boss/GPT-SoVITS/pull/1267) 노이즈 제거 또는 ASR이 예외를 만나면 모든 보류 중인 오디오 파일이 종료되는 문제를 수정했습니다.
+3. [PR 1253](https://github.com/RVC-Boss/GPT-SoVITS/pull/1253) 구두점으로 분할할 때 소수점 분할 문제를 수정했습니다.
+4. [Commit a208698](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca) 다중 GPU 훈련을 위한 다중 프로세스 저장 로직을 수정했습니다.
+5. [PR 1251](https://github.com/RVC-Boss/GPT-SoVITS/pull/1251) 불필요한 `my_utils`를 제거했습니다.
+
+#### 주요 수정:
+
+6. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672)의 가속 추론 코드가 검증되어 메인 브랜치에 병합되었으며, 기본과 일관된 추론 효과를 보장합니다.
+   또한 참조 텍스트 없는 모드에서 가속 추론을 지원합니다.
+
+**향후 업데이트에서는 `fast_inference` 브랜치의 변경 사항의 일관성을 계속 검증할 것입니다**.
+
+### 20240727 업데이트
+
+#### 사소한 수정:
+
+1. [PR 1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298) 불필요한 i18n 코드를 정리했습니다.
+2. [PR 1299](https://github.com/RVC-Boss/GPT-SoVITS/pull/1299) 사용자 파일 경로의 후행 슬래시가 명령줄 오류를 일으키는 문제를 수정했습니다.
+3. [PR 756](https://github.com/RVC-Boss/GPT-SoVITS/pull/756) GPT 훈련의 단계 계산 로직을 수정했습니다.
+
+#### 주요 수정:
+
+4. [Commit 9588a3c](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2) 합성을 위한 음성 속도 조절을 지원했습니다.
+   음성 속도만 조절하면서 무작위성을 고정할 수 있습니다.
+
+### 20240806 업데이트
+
+1. [PR 1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR 1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) BS RoFormer 보컬 반주 분리 모델에 대한 지원을 추가했습니다. [Commit e62e965](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c) FP16 추론을 활성화했습니다.
+2. 중국어 텍스트 프론트엔드를 개선했습니다.
+   - [PR 488](https://github.com/RVC-Boss/GPT-SoVITS/pull/488) 중국어 다의자 지원 (v2 전용);
+   - [PR 987](https://github.com/RVC-Boss/GPT-SoVITS/pull/987) 추가된 양자;
+   - [PR 1351](https://github.com/RVC-Boss/GPT-SoVITS/pull/1351) 사칙연산 및 기본 수학 공식을 지원합니다;
+   - [PR 1404](https://github.com/RVC-Boss/GPT-SoVITS/pull/1404) 혼합 텍스트 오류를 수정했습니다.
+3. [PR 1355](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) WebUI 에서 오디오를 처리할 때 경로를 자동으로 채웠습니다.
+4. [Commit bce451a](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299), [Commit 4c8b761](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78) GPU 인식 로직을 최적화했습니다.
+5. [Commit 8a10147](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3) 광동어 ASR 지원을 추가했습니다.
+6. GPT-SoVITS v2 지원을 추가했습니다.
+7. [PR 1387](https://github.com/RVC-Boss/GPT-SoVITS/pull/1387) 타이밍 로직을 최적화했습니다.
+
+### 20240821 업데이트
+
+1. [PR 1490](https://github.com/RVC-Boss/GPT-SoVITS/pull/1490) `fast_inference` 브랜치를 메인 브랜치에 병합.
+2. [Issue 1508](https://github.com/RVC-Boss/GPT-SoVITS/issues/1508) SSML 태그를 사용하여 숫자, 전화번호, 날짜 및 시간 최적화 지원.
+3. [PR 1503](https://github.com/RVC-Boss/GPT-SoVITS/pull/1503) API 수정 및 최적화.
+4. [PR 1422](https://github.com/RVC-Boss/GPT-SoVITS/pull/1422) 믹싱을 위한 참조 오디오를 하나만 업로드할 수 있는 버그 수정, 다양한 데이터셋 검사 추가 및 파일이 누락된 경우 경고 팝업.
+
+### 20250211 업데이트
+
+- [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)) GPT-SoVITS v3 모델 추가, SoVITS v3의 파인튜닝에는 14GB GPU 메모리가 필요합니다.
+
+### 20250212 업데이트
+
+- [PR 2040](https://github.com/RVC-Boss/GPT-SoVITS/pull/2040) SoVITS v3의 파인튜닝에 그라디언트 체크포인트 추가, 12GB GPU 메모리가 필요합니다.
+
+### 20250214 업데이트
+
+- [PR 2047](https://github.com/RVC-Boss/GPT-SoVITS/pull/2047) 다국어 혼합 텍스트 분할 전략 **A** 최적화.
+  - `split-lang`을 언어 분할 도구로 추가하여 다국어 혼합 텍스트의 분할 능력을 향상시켰습니다.
+
+### 20250217 업데이트
+
+- [PR 2062](https://github.com/RVC-Boss/GPT-SoVITS/pull/2062) 텍스트 내 숫자와 영어 처리 로직 최적화.
+
+### 20250218 업데이트
+
+- [PR 2073](https://github.com/RVC-Boss/GPT-SoVITS/pull/2073) 다국어 혼합 텍스트 분할 전략 **B** 최적화.
+
+### 20250223 업데이트
+
+1. SoVITS V3의 파인튜닝에 LoRA 훈련이 지원됩니다. 8GB GPU 메모리가 필요하며, 전체 매개변수 파인튜닝보다 더 나은 결과를 제공합니다.
+2. [PR 2078](https://github.com/RVC-Boss/GPT-SoVITS/pull/2078) 보컬 및 악기 분리를 위해 Mel Band RoFormer 모델 추가.
+
+### 20250226 업데이트
+
+1. [PR 2112](https://github.com/RVC-Boss/GPT-SoVITS/pull/2112) Windows에서 비영어 디렉토리로 인한 문제 수정.
+   - 한국어에 대한 `langsegmenter` 사용 문제 수정.
+2. [PR 2113](https://github.com/RVC-Boss/GPT-SoVITS/pull/2114) Windows에서 비영어 디렉토리로 인한 문제 수정.
+   - 한국어/일본어에 대한 `langsegmenter` 사용 문제 수정.
+
+### 20250227 업데이트
+
+- V3 모델로 24K 오디오를 생성할 때 발생하는 음성 뭉침 문제를 완화하기 위해, 24K에서 48K로의 오디오 초해상도 모델을 추가했습니다. [Issue 2085](https://github.com/RVC-Boss/GPT-SoVITS/issues/2085), [Issue 2117](https://github.com/RVC-Boss/GPT-SoVITS/issues/2117)에서 보고된 문제입니다.
\ No newline at end of file
diff --git a/docs/ko/README.md b/docs/ko/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..e9fee8fc99592ffb9c9bb47a1482a92ded64c093
--- /dev/null
+++ b/docs/ko/README.md
@@ -0,0 +1,389 @@
+<div align="center">
+
+<h1>GPT-SoVITS-WebUI</h1>
+소량의 데이터로 음성 변환 및 음성 합성을 지원하는 강력한 WebUI.<br><br>
+
+[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
+
+<img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br>
+
+[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
+[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
+[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
+[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
+
+[**English**](../../README.md) | [**中文简体**](../cn/README.md) | [**日本語**](../ja/README.md) | **한국어** | [**Türkçe**](../tr/README.md)
+
+</div>
+
+---
+
+## 기능:
+
+1. **제로샷 텍스트 음성 변환 (TTS):** 5초의 음성 샘플을 입력하면 즉시 텍스트를 음성으로 변환할 수 있습니다.
+
+2. **소량의 데이터 TTS:** 1분의 훈련 데이터만으로 모델을 미세 조정하여 음성 유사도와 실제감을 향상시킬 수 있습니다.
+
+3. **다국어 지원:** 훈련 데이터셋과 다른 언어의 추론을 지원하며, 현재 영어, 일본어, 중국어, 광둥어, 한국어를 지원합니다.
+
+4. **WebUI 도구:** 음성 반주 분리, 자동 훈련 데이터셋 분할, 중국어 자동 음성 인식(ASR) 및 텍스트 주석 등의 도구를 통합하여 초보자가 훈련 데이터셋과 GPT/SoVITS 모델을 생성하는 데 도움을 줍니다.
+
+**데모 비디오를 확인하세요! [demo video](https://www.bilibili.com/video/BV12g4y1m7Uw)**
+
+보지 못한 발화자의 퓨샷(few-shot) 파인튜닝 데모:
+
+https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
+
+**사용자 설명서: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
+
+## 설치
+
+### 테스트 통과 환경
+
+| Python Version | PyTorch Version  | Device          |
+|----------------|------------------|-----------------|
+| Python 3.9     | PyTorch 2.0.1    | CUDA 11.8       |
+| Python 3.10.13 | PyTorch 2.1.2    | CUDA 12.3       |
+| Python 3.10.17 | PyTorch 2.5.1    | CUDA 12.4       |
+| Python 3.9     | PyTorch 2.5.1    | Apple silicon   |
+| Python 3.11    | PyTorch 2.6.0    | Apple silicon   |
+| Python 3.9     | PyTorch 2.2.2    | CPU             |
+
+### Windows
+
+Windows 사용자라면 (win>=10에서 테스트됨), [통합 패키지를 다운로드](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-v3lora-20250228.7z?download=true)한 후 압축을 풀고 _go-webui.bat_ 파일을 더블 클릭하면 GPT-SoVITS-WebUI를 시작할 수 있습니다.
+
+### Linux
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### macOS
+
+**주의: Mac에서 GPU로 훈련된 모델은 다른 OS에서 훈련된 모델에 비해 품질이 낮습니다. 해당 문제를 해결하기 전까지 MacOS에선 CPU를 사용하여 훈련을 진행합니다.**
+
+1. `xcode-select --install`을 실행하여 Xcode 커맨드라인 도구를 설치하세요.
+2. 다음 명령어를 실행하여 이 프로젝트를 설치하세요.
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### 수동 설치
+
+#### FFmpeg 설치
+
+##### Conda 사용자
+
+```bash
+conda install ffmpeg
+```
+
+##### Ubuntu/Debian 사용자
+
+```bash
+sudo apt install ffmpeg
+sudo apt install libsox-dev
+conda install -c conda-forge 'ffmpeg<7'
+```
+
+##### Windows 사용자
+
+[ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe)와 [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe)를 GPT-SoVITS root 디렉토리에 넣습니다.
+
+[Visual Studio 2017](https://aka.ms/vs/17/release/vc_redist.x86.exe) 설치 (Korean TTS 전용)
+
+##### MacOS 사용자
+
+```bash
+brew install ffmpeg
+```
+
+#### 의존성 설치
+
+```bash
+pip install -r extra-req.txt --no-deps
+pip install -r requirements.txt
+```
+
+### Docker에서 사용
+
+#### docker-compose.yaml 설정
+
+0. 이미지 태그: 코드 저장소가 빠르게 업데이트되고 패키지가 느리게 빌드되고 테스트되므로, 현재 빌드된 최신 도커 이미지를 [Docker Hub](https://hub.docker.com/r/breakstring/gpt-sovits)(오래된 버전) 에서 확인하고 필요에 따라 Dockerfile을 사용하여 로컬에서 빌드할 수 있습니다.
+
+1. 환경 변수:
+
+- is_half: 반정밀/배정밀 제어. "SSL 추출" 단계에서 4-cnhubert/5-wav32k 디렉토리의 내용을 올바르게 생성할 수 없는 경우, 일반적으로 이것 때문입니다. 실제 상황에 따라 True 또는 False로 조정할 수 있습니다.
+
+2. 볼륨 설정, 컨테이너 내의 애플리케이션 루트 디렉토리를 /workspace로 설정합니다. 기본 docker-compose.yaml에는 실제 예제가 나열되어 있으므로 업로드/다운로드를 쉽게 할 수 있습니다.
+
+3. shm_size: Windows의 Docker Desktop의 기본 사용 가능한 메모리가 너무 작아 오류가 발생할 수 있으므로 실제 상황에 따라 조정합니다.
+
+4. deploy 섹션의 gpu 관련 내용은 시스템 및 실제 상황에 따라 조정합니다.
+
+#### docker compose로 실행
+
+```
+docker compose -f "docker-compose.yaml" up -d
+```
+
+#### docker 명령으로 실행
+
+위와 동일하게 실제 상황에 맞게 매개변수를 수정한 다음 다음 명령을 실행합니다:
+
+```
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+```
+
+## 사전 학습된 모델
+
+**`install.sh`가 성공적으로 실행되면 No.1,2,3 은 건너뛰어도 됩니다.**
+
+1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) 에서 사전 학습된 모델을 다운로드하고, `GPT_SoVITS/pretrained_models` 디렉토리에 배치하세요.
+
+2. [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) 에서 모델을 다운로드하고 압축을 풀어 `G2PWModel`로 이름을 변경한 후, `GPT_SoVITS/text` 디렉토리에 배치하세요. (중국어 TTS 전용)
+
+3. UVR5 (보컬/반주 분리 & 잔향 제거 추가 기능)의 경우, [UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) 에서 모델을 다운로드하고 `tools/uvr5/uvr5_weights` 디렉토리에 배치하세요.
+
+   - UVR5에서 bs_roformer 또는 mel_band_roformer 모델을 사용할 경우, 모델과 해당 설정 파일을 수동으로 다운로드하여 `tools/UVR5/UVR5_weights` 폴더에 저장할 수 있습니다. **모델 파일과 설정 파일의 이름은 확장자를 제외하고 동일한 이름을 가지도록 해야 합니다**. 또한, 모델과 설정 파일 이름에는 **"roformer"**가 포함되어야 roformer 클래스의 모델로 인식됩니다.
+
+   - 모델 이름과 설정 파일 이름에 **모델 유형을 직접 지정하는 것이 좋습니다**. 예: mel_mand_roformer, bs_roformer. 지정하지 않으면 설정 파일을 기준으로 특성을 비교하여 어떤 유형의 모델인지를 판단합니다. 예를 들어, 모델 `bs_roformer_ep_368_sdr_12.9628.ckpt`와 해당 설정 파일 `bs_roformer_ep_368_sdr_12.9628.yaml`은 한 쌍입니다. `kim_mel_band_roformer.ckpt`와 `kim_mel_band_roformer.yaml`도 한 쌍입니다.
+
+4. 중국어 ASR (추가 기능)의 경우, [Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files), [Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files) 및 [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) 에서 모델을 다운로드하고, `tools/asr/models` 디렉토리에 배치하세요.
+
+5. 영어 또는 일본어 ASR (추가 기능)의 경우, [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) 에서 모델을 다운로드하고, `tools/asr/models` 디렉토리에 배치하세요. 또한, [다른 모델](https://huggingface.co/Systran) 은 더 적은 디스크 용량으로 비슷한 효과를 가질 수 있습니다.
+
+## 데이터셋 형식
+
+텍스트 음성 합성(TTS) 주석 .list 파일 형식:
+
+```
+vocal_path|speaker_name|language|text
+```
+
+언어 사전:
+
+- 'zh': 중국어
+- 'ja': 일본어
+- 'en': 영어
+
+예시:
+
+```
+D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
+```
+
+## 미세 조정 및 추론
+
+### WebUI 열기
+
+#### 통합 패키지 사용자
+
+`go-webui.bat`을 더블 클릭하거나 `go-webui.ps1`를 사용하십시오.
+V1으로 전환하려면, `go-webui-v1.bat`을 더블 클릭하거나 `go-webui-v1.ps1`를 사용하십시오.
+
+#### 기타
+
+```bash
+python webui.py <언어(옵션)>
+```
+
+V1으로 전환하려면,
+
+```bash
+python webui.py v1 <언어(옵션)>
+```
+
+또는 WebUI에서 수동으로 버전을 전환하십시오.
+
+### 미세 조정
+
+#### 경로 자동 채우기가 지원됩니다
+
+    1. 오디오 경로를 입력하십시오.
+    2. 오디오를 작은 청크로 분할하십시오.
+    3. 노이즈 제거(옵션)
+    4. ASR 수행
+    5. ASR 전사를 교정하십시오.
+    6. 다음 탭으로 이동하여 모델을 미세 조정하십시오.
+
+### 추론 WebUI 열기
+
+#### 통합 패키지 사용자
+
+`go-webui-v2.bat`을 더블 클릭하거나 `go-webui-v2.ps1`를 사용한 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다.
+
+#### 기타
+
+```bash
+python GPT_SoVITS/inference_webui.py <언어(옵션)>
+```
+
+또는
+
+```bash
+python webui.py
+```
+
+그런 다음 `1-GPT-SoVITS-TTS/1C-inference`에서 추론 webui를 엽니다.
+
+## V2 릴리스 노트
+
+새로운 기능:
+
+1. 한국어 및 광둥어 지원
+
+2. 최적화된 텍스트 프론트엔드
+
+3. 사전 학습 모델이 2천 시간에서 5천 시간으로 확장
+
+4. 저품질 참조 오디오에 대한 합성 품질 향상
+
+   [자세한 내용](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v2%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+V1 환경에서 V2를 사용하려면:
+
+1. `pip install -r requirements.txt`를 사용하여 일부 패키지 업데이트
+
+2. github에서 최신 코드를 클론하십시오.
+
+3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained)에서 V2 사전 학습 모델을 다운로드하여 `GPT_SoVITS\pretrained_models\gsv-v2final-pretrained`에 넣으십시오.
+
+   중국어 V2 추가: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW 모델을 다운로드하여 압축을 풀고 `G2PWModel`로 이름을 변경한 다음 `GPT_SoVITS/text`에 배치합니다.)
+
+## V3 릴리스 노트
+
+새로운 기능:
+
+1. 음색 유사성이 더 높아져 목표 음성에 대한 학습 데이터가 적게 필요합니다. (기본 모델을 직접 사용하여 미세 조정 없이 음색 유사성이 크게 향상됩니다.)
+
+2. GPT 모델이 더 안정적이며 반복 및 생략이 적고, 더 풍부한 감정 표현을 가진 음성을 생성하기가 더 쉽습니다.
+
+   [자세한 내용](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+v2 환경에서 v3 사용하기:
+
+1. `pip install -r requirements.txt`로 일부 패키지를 업데이트합니다.
+
+2. 최신 코드를 github 에서 클론합니다.
+
+3. v3 사전 훈련된 모델(s1v3.ckpt, s2Gv3.pth, 그리고 models--nvidia--bigvgan_v2_24khz_100band_256x 폴더)을 [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main)에서 다운로드하여 `GPT_SoVITS\pretrained_models` 폴더에 넣습니다.
+
+   추가: 오디오 슈퍼 해상도 모델에 대해서는 [다운로드 방법](../../tools/AP_BWE_main/24kto48k/readme.txt)을 참고하세요.
+
+## 할 일 목록
+
+- [x] **최우선순위:**
+
+  - [x] 일본어 및 영어 지역화.
+  - [x] 사용자 가이드.
+  - [x] 일본어 및 영어 데이터셋 미세 조정 훈련.
+
+- [ ] **기능:**
+
+  - [x] 제로샷 음성 변환 (5초) / 소량의 음성 변환 (1분).
+  - [x] TTS 속도 제어.
+  - [ ] ~~향상된 TTS 감정 제어.~~
+  - [ ] SoVITS 토큰 입력을 단어 확률 분포로 변경해 보세요.
+  - [x] 영어 및 일본어 텍스트 프론트 엔드 개선.
+  - [ ] 작은 크기와 큰 크기의 TTS 모델 개발.
+  - [x] Colab 스크립트.
+  - [ ] 훈련 데이터셋 확장 (2k 시간에서 10k 시간).
+  - [x] 더 나은 sovits 기본 모델 (향상된 오디오 품질).
+  - [ ] 모델 블렌딩.
+
+## (추가적인) 명령줄에서 실행하는 방법
+
+명령줄을 사용하여 UVR5용 WebUI 열기
+
+```
+python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
+```
+
+<!-- 브라우저를 열 수 없는 경우 UVR 처리를 위해 아래 형식을 따르십시오. 이는 오디오 처리를 위해 mdxnet을 사용하는 것입니다.
+```
+python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
+``` -->
+
+명령줄을 사용하여 데이터세트의 오디오 분할을 수행하는 방법은 다음과 같습니다.
+
+```
+python audio_slicer.py \
+    --input_path "<path_to_original_audio_file_or_directory>" \
+    --output_root "<directory_where_subdivided_audio_clips_will_be_saved>" \
+    --threshold <volume_threshold> \
+    --min_length <minimum_duration_of_each_subclip> \
+    --min_interval <shortest_time_gap_between_adjacent_subclips>
+    --hop_size <step_size_for_computing_volume_curve>
+```
+
+명령줄을 사용하여 데이터 세트 ASR 처리를 수행하는 방법입니다(중국어만 해당).
+
+```
+python tools/asr/funasr_asr.py -i <input> -o <output>
+```
+
+ASR 처리는 Faster_Whisper(중국어를 제외한 ASR 마킹)를 통해 수행됩니다.
+
+(진행률 표시줄 없음, GPU 성능으로 인해 시간 지연이 발생할 수 있음)
+
+```
+python ./tools/asr/fasterwhisper_asr.py -i <input> -o <output> -l <language> -p <precision>
+```
+
+사용자 정의 목록 저장 경로가 활성화되었습니다.
+
+## 감사의 말
+
+다음 프로젝트와 기여자들에게 특별히 감사드립니다:
+
+### 이론 연구
+
+- [ar-vits](https://github.com/innnky/ar-vits)
+- [SoundStorm](https://github.com/yangdongchao/SoundStorm/tree/master/soundstorm/s1/AR)
+- [vits](https://github.com/jaywalnut310/vits)
+- [TransferTTS](https://github.com/hcy71o/TransferTTS/blob/master/models.py#L556)
+- [contentvec](https://github.com/auspicious3000/contentvec/)
+- [hifi-gan](https://github.com/jik876/hifi-gan)
+- [fish-speech](https://github.com/fishaudio/fish-speech/blob/main/tools/llama/generate.py#L41)
+- [f5-TTS](https://github.com/SWivid/F5-TTS/blob/main/src/f5_tts/model/backbones/dit.py)
+- [shortcut flow matching](https://github.com/kvfrans/shortcut-models/blob/main/targets_shortcut.py)
+
+### 사전 학습 모델
+
+- [Chinese Speech Pretrain](https://github.com/TencentGameMate/chinese_speech_pretrain)
+- [Chinese-Roberta-WWM-Ext-Large](https://huggingface.co/hfl/chinese-roberta-wwm-ext-large)
+- [BigVGAN](https://github.com/NVIDIA/BigVGAN)
+
+### 추론용 텍스트 프론트엔드
+
+- [paddlespeech zh_normalization](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization)
+- [split-lang](https://github.com/DoodleBears/split-lang)
+- [g2pW](https://github.com/GitYCC/g2pW)
+- [pypinyin-g2pW](https://github.com/mozillazg/pypinyin-g2pW)
+- [paddlespeech g2pw](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/g2pw)
+
+### WebUI 도구
+
+- [ultimatevocalremovergui](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [SubFix](https://github.com/cronrpc/SubFix)
+- [FFmpeg](https://github.com/FFmpeg/FFmpeg)
+- [gradio](https://github.com/gradio-app/gradio)
+- [faster-whisper](https://github.com/SYSTRAN/faster-whisper)
+- [FunASR](https://github.com/alibaba-damo-academy/FunASR)
+- [AP-BWE](https://github.com/yxlu-0102/AP-BWE)
+
+@Naozumi520 님께 감사드립니다. 광둥어 학습 자료를 제공해 주시고, 광둥어 관련 지식을 지도해 주셔서 감사합니다.
+
+## 모든 기여자들에게 감사드립니다 ;)
+
+<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
+  <img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
+</a>
diff --git a/docs/tr/Changelog_TR.md b/docs/tr/Changelog_TR.md
new file mode 100644
index 0000000000000000000000000000000000000000..3de7c8a2a4f1d152f52fdd3b59f57aa365b09da7
--- /dev/null
+++ b/docs/tr/Changelog_TR.md
@@ -0,0 +1,222 @@
+### 20240121 Güncellemesi
+
+1. `config`e `is_share` eklendi. Colab gibi senaryolarda, WebUI'yi halka açık ağa yönlendirmek için bu değeri `True` olarak ayarlayabilirsiniz.
+2. WebUI'ye İngilizce sistem çeviri desteği eklendi.
+3. `cmd-asr`, FunASR modelinin dahil olup olmadığını otomatik olarak tespit eder; eğer varsayılan dizinde bulunamazsa, ModelScope'dan indirilecektir.
+4. [Issue 79](https://github.com/RVC-Boss/GPT-SoVITS/issues/79)de bildirilen SoVITS eğitimindeki ZeroDivisionError'u sıfır uzunlukta örnekleri filtreleyerek düzeltmeye çalıştık.
+5. `TEMP` klasöründeki önbelleğe alınmış ses dosyaları ve diğer dosyaları temizledik.
+6. Referans sesinin sonunu içeren sentezlenmiş ses sorununu önemli ölçüde azalttık.
+
+### 20240122 Güncellemesi
+
+1. Aşırı kısa çıktı dosyalarının referans sesini tekrarlamasına neden olan sorun giderildi.
+2. İngilizce ve Japonca eğitim için yerel destek test edildi (Japonca eğitim için kök dizinin İngilizce olmayan özel karakterlerden arındırılmış olması gerekir).
+3. Ses yolu denetimi iyileştirildi. Yanlış bir giriş yolundan okumaya çalışıldığında, ffmpeg hatası yerine yolun mevcut olmadığını bildirir.
+
+### 20240123 Güncellemesi
+
+1. Hubert çıkarımının NaN hatalarına neden olup SoVITS/GPT eğitiminde ZeroDivisionError'a yol açtığı sorun çözüldü.
+2. İnferans WebUI'de hızlı model değiştirme desteği eklendi.
+3. Model dosyası sıralama mantığı optimize edildi.
+4. Çince kelime ayrımı için `jieba` `jieba_fast` ile değiştirildi.
+
+### 20240126 Güncellemesi
+
+1. Çince-İngilizce ve Japonca-İngilizce karışık çıktı metinleri için destek eklendi.
+2. Çıktı için isteğe bağlı bir bölme modu eklendi.
+3. UVR5'in dizinlerden otomatik olarak çıkmasına neden olan okuma sorununu düzelttik.
+4. Çeşitli yeni satır sorunlarını düzelterek çıkarım hatalarını giderdik.
+5. Çıkarım WebUI'deki gereksiz günlükleri kaldırdık.
+6. Mac'te eğitim ve çıkarım desteği eklendi.
+7. Yarım hassasiyeti desteklemeyen GPU'lar için otomatik olarak tek hassasiyet zorlandı; CPU çıkarımında tek hassasiyet uygulandı.
+
+### 20240128 Güncellemesi
+
+1. Sayıların Çince karakterlere dönüştürülmesiyle ilgili sorunu düzelttik.
+2. Cümlelerin başındaki birkaç karakterin yutulması sorununu düzelttik.
+3. Mantıksız referans ses uzunluklarını sınırlamalar koyarak hariç tuttuk.
+4. GPT eğitiminin kontrol noktalarını kaydetmemesi sorununu düzelttik.
+5. Dockerfile'da model indirme sürecini tamamladık.
+
+### 20240129 Güncellemesi
+
+1. Yarım hassasiyet eğitimi ile ilgili sorun yaşayan 16 serisi gibi GPU'lar için eğitim yapılandırmalarını tek hassasiyete değiştirdik.
+2. Mevcut Colab sürümünü test ettik ve güncelledik.
+3. Eski sürüm FunASR ile ModelScope FunASR deposunun git klonlanmasıyla oluşan arayüz hizalama hatalarını düzelttik.
+
+### 20240130 Güncellemesi
+
+1. Çift tırnaklarla yol kopyalama hatalarını önlemek için tüm yol ile ilgili girdilerden otomatik olarak çift tırnakları kaldırdık.
+2. Çince ve İngilizce noktalama işaretlerini ayırma sorunlarını düzelttik ve cümlelerin başına ve sonuna noktalama işaretleri ekledik.
+3. Noktalama işaretlerine göre ayırma özelliğini ekledik.
+
+### 20240201 Güncellemesi
+
+1. Ayrılma hatalarına neden olan UVR5 format okuma hatasını düzelttik.
+2. Karışık Çince-Japonca-İngilizce metinler için otomatik segmentasyon ve dil tanıma desteği sağladık.
+
+### 20240202 Güncellemesi
+
+1. `/` ile biten bir ASR yolunun dosya adını kaydetme hatasına neden olma sorununu düzelttik.
+2. [PR 377](https://github.com/RVC-Boss/GPT-SoVITS/pull/377) PaddleSpeech'in Normalizer'ını tanıtarak "xx.xx%" (yüzde sembolleri) ve "元/吨" ifadesinin "元吨" yerine "元每吨" olarak okunması gibi sorunları düzelttik ve alt çizgi hatalarını giderdik.
+
+### 20240207 Güncellemesi
+
+1. [Issue 391](https://github.com/RVC-Boss/GPT-SoVITS/issues/391)de bildirilen dil parametresi karışıklığının Çinçe çıkarım kalitesini düşürme sorununu düzelttik.
+2. [PR 403](https://github.com/RVC-Boss/GPT-SoVITS/pull/403) ile UVR5'i daha yüksek versiyonlarda librosa'ya uyarladık.
+3. [Commit 14a2851](https://github.com/RVC-Boss/GPT-SoVITS/commit/14a285109a521679f8846589c22da8f656a46ad8) `is_half` parametresinin booleana dönüştürülmemesi nedeniyle sürekli yarım hassasiyet çıkarımı yaparak 16 serisi GPU'larda `inf` hatasına neden olan UVR5 inf hatasını düzelttik.
+4. İngilizce metin önyüzünü optimize ettik.
+5. Gradio bağımlılıklarını düzelttik.
+6. Veri seti hazırlığı sırasında kök dizini boş bırakıldığında `.list` tam yollarının otomatik olarak okunmasını destekledik.
+7. Japonca ve İngilizce için Faster Whisper ASR'yi entegre ettik.
+
+### 20240208 Güncellemesi
+
+1. [Commit 59f35ad](https://github.com/RVC-Boss/GPT-SoVITS/commit/59f35adad85815df27e9c6b33d420f5ebfd8376b) ile Windows 10 1909'da ve [Issue 232](https://github.com/RVC-Boss/GPT-SoVITS/issues/232)de (Geleneksel Çince Sistem Dili) bildirilen GPT eğitim durma sorununu düzeltmeye çalıştık.
+
+### 20240212 Güncellemesi
+
+1. Faster Whisper ve FunASR için mantığı optimize ettik, Hugging Face bağlantı sorunlarını önlemek için Faster Whisper'ı ayna indirmelere yönlendirdik.
+2. [PR 457](https://github.com/RVC-Boss/GPT-SoVITS/pull/457) GPT tekrarı ve eksik karakterleri azaltmak için eğitim sırasında negatif örnekler oluşturarak deneysel DPO Loss eğitim seçeneğini etkinleştirdi ve çıkarım WebUI'de çeşitli çıkarım parametrelerini kullanılabilir hale getirdi.
+
+### 20240214 Güncellemesi
+
+1. Eğitimde Çince deney adlarını destekledik (önceden hatalara neden oluyordu).
+2. DPO eğitimini zorunlu yerine isteğe bağlı bir özellik yaptık. Seçilirse, parti boyutu otomatik olarak yarıya indirilir. Çıkarım WebUI'de yeni parametrelerin iletilmemesi sorunlarını düzelttik.
+
+### 20240216 Güncellemesi
+
+1. Referans metin olmadan girişi destekledik.
+2. [Issue 475](https://github.com/RVC-Boss/GPT-SoVITS/issues/475) de bildirilen Çince önyüz hatalarını düzelttik.
+
+### 20240221 Güncellemesi
+
+1. Veri işleme sırasında bir gürültü azaltma seçeneği ekledik (gürültü azaltma sadece 16kHz örnekleme hızını bırakır; yalnızca arka plan gürültüsü önemliyse kullanın).
+2. [PR 559](https://github.com/RVC-Boss/GPT-SoVITS/pull/559), [PR 556](https://github.com/RVC-Boss/GPT-SoVITS/pull/556), [PR 532](https://github.com/RVC-Boss/GPT-SoVITS/pull/532), [PR 507](https://github.com/RVC-Boss/GPT-SoVITS/pull/507), [PR 509](https://github.com/RVC-Boss/GPT-SoVITS/pull/509) ile Çince ve Japonca önyüz işlemesini optimize ettik.
+3. Mac CPU çıkarımını daha hızlı performans için MPS yerine CPU kullanacak şekilde değiştirdik.
+4. Colab genel URL sorununu düzelttik.
+
+### 20240306 Güncellemesi
+
+1. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) çıkarımı %50 hızlandırdı (RTX3090 + PyTorch 2.2.1 + CU11.8 + Win10 + Py39 üzerinde test edildi).
+2. Faster Whisper'ın Çince olmayan ASR'sini kullanırken artık önce Çin FunASR modelini indirmeyi gerektirmiyor.
+3. [PR 610](https://github.com/RVC-Boss/GPT-SoVITS/pull/610) UVR5 yankı giderme modelindeki ayarın tersine çevrildiği sorunu düzeltti.
+4. [PR 675](https://github.com/RVC-Boss/GPT-SoVITS/pull/675) CUDA mevcut olmadığında Faster Whisper için otomatik CPU çıkarımını etkinleştirdi.
+5. [PR 573](https://github.com/RVC-Boss/GPT-SoVITS/pull/573) Mac'te doğru CPU çıkarımı sağlamak için `is_half` kontrolünü değiştirdi.
+
+### 202403/202404/202405 Güncellemeleri
+
+#### Küçük Düzeltmeler:
+
+1. Referans metin olmayan mod ile ilgili sorunlar düzeltildi.
+2. Çince ve İngilizce metin önyüzü optimize edildi.
+3. API formatı iyileştirildi.
+4. CMD format sorunları düzeltildi.
+5. Eğitim verisi işleme sırasında desteklenmeyen diller için hata uyarıları eklendi.
+6. Hubert çıkarımındaki hata düzeltildi.
+
+#### Büyük Düzeltmeler:
+
+1. VQ'yu dondurmadan yapılan SoVITS eğitimi sorunu (bu kalite düşüşüne neden olabilir) düzeltildi.
+2. Hızlı çıkarım dalı eklendi.
+
+### 20240610 Güncellemesi
+
+#### Küçük Düzeltmeler:
+
+1. [PR 1168](https://github.com/RVC-Boss/GPT-SoVITS/pull/1168) & [PR 1169](https://github.com/RVC-Boss/GPT-SoVITS/pull/1169) saf noktalama işareti ve çoklu noktalama işareti metin girdisi için mantığı geliştirdi.
+2. [Commit 501a74a](https://github.com/RVC-Boss/GPT-SoVITS/commit/501a74ae96789a26b48932babed5eb4e9483a232) UVR5'teki MDXNet yankı giderme için CMD formatını düzeltti, boşluk içeren yolları destekledi.
+3. [PR 1159](https://github.com/RVC-Boss/GPT-SoVITS/pull/1159) `s2_train.py` içindeki SoVITS eğitimi için ilerleme çubuğu mantığını düzeltti.
+
+#### Büyük Düzeltmeler:
+
+4. [Commit 99f09c8](https://github.com/RVC-Boss/GPT-SoVITS/commit/99f09c8bdc155c1f4272b511940717705509582a) WebUI'nin GPT ince ayarının, Çince giriş metinlerinin BERT özelliğini okumaması sorununu düzeltti, bu da çıkarım ile tutarsızlığa ve potansiyel kalite düşüşüne neden oluyordu.
+   **Dikkat: Daha önce büyük miktarda veri ile ince ayar yaptıysanız, modelin kalitesini artırmak için yeniden ayar yapmanız önerilir.**
+
+### 20240706 Güncellemesi
+
+#### Küçük Düzeltmeler:
+
+1. [Commit 1250670](https://github.com/RVC-Boss/GPT-SoVITS/commit/db50670598f0236613eefa6f2d5a23a271d82041) CPU çıkarımında varsayılan yığın boyutu ondalık sorununu düzeltti.
+2. [PR 1258](https://github.com/RVC-Boss/GPT-SoVITS/pull/1258), [PR 1265](https://github.com/RVC-Boss/GPT-SoVITS/pull/1265), [PR 1267](https://github.com/RVC-Boss/GPT-SoVITS/pull/1267) gürültü giderme veya ASR ile ilgili istisnalarla karşılaşıldığında bekleyen tüm ses dosyalarının çıkış yapmasına neden olan sorunları düzeltti.
+3. [PR 1253](https://github.com/RVC-Boss/GPT-SoVITS/pull/1253) noktalama işaretlerine göre ayrılırken ondalıkların bölünmesi sorununu düzeltti.
+4. [Commit a208698](https://github.com/RVC-Boss/GPT-SoVITS/commit/a208698e775155efc95b187b746d153d0f2847ca) çoklu GPU eğitimi için çoklu işlem kaydetme mantığını düzeltti.
+5. [PR 1251](https://github.com/RVC-Boss/GPT-SoVITS/pull/1251) gereksiz `my_utils`'ı kaldırdı.
+
+#### Büyük Düzeltmeler:
+
+6. [PR 672](https://github.com/RVC-Boss/GPT-SoVITS/pull/672) hızlandırılmış çıkarım kodu doğrulandı ve ana dala birleştirildi, taban ile tutarlı çıkarım etkileri sağlandı.
+   Ayrıca referans metni olmayan modda hızlandırılmış çıkarımı destekler.
+
+**Gelecek güncellemeler, `fast_inference` dalındaki değişikliklerin tutarlılığını doğrulamaya devam edecek.**
+
+### 20240727 Güncellemesi
+
+#### Küçük Düzeltmeler:
+
+1. [PR 1298](https://github.com/RVC-Boss/GPT-SoVITS/pull/1298) gereksiz i18n kodlarını temizledi.
+2. [PR 1299](https://github.com/RVC-Boss/GPT-SoVITS/pull/1299) kullanıcı dosya yollarındaki sonlandırma eğik çizgilerinin komut satırı hatalarına neden olduğu sorunları düzeltti.
+3. [PR 756](https://github.com/RVC-Boss/GPT-SoVITS/pull/756) GPT eğitimindeki adım hesaplama mantığını düzeltti.
+
+#### Büyük Düzeltmeler:
+
+4. [Commit 9588a3c](https://github.com/RVC-Boss/GPT-SoVITS/commit/9588a3c52d9ebdb20b3c5d74f647d12e7c1171c2) sentez için konuşma hızı ayarlamasını destekledi.
+   Konuşma hızını ayarlarken rastgeleliği dondurmayı etkinleştirdi.
+
+### 20240806 Güncellemesi
+
+1. [PR 1306](https://github.com/RVC-Boss/GPT-SoVITS/pull/1306), [PR 1356](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) BS RoFormer vokal eşlik ayırma modelini desteklemeye başladı. [Commit e62e965](https://github.com/RVC-Boss/GPT-SoVITS/commit/e62e965323a60a76a025bcaa45268c1ddcbcf05c) FP16 çıkarımı etkinleştirdi.
+2. Çince metin ön yüzünü geliştirdi.
+   - [PR 488](https://github.com/RVC-Boss/GPT-SoVITS/pull/488) çoklu heceli karakterler için destek ekledi (v2 sadece);
+   - [PR 987](https://github.com/RVC-Boss/GPT-SoVITS/pull/987) sayı belirleyici ekledi;
+   - [PR 1351](https://github.com/RVC-Boss/GPT-SoVITS/pull/1351) aritmetik ve temel matematik formüllerini destekler;
+   - [PR 1404](https://github.com/RVC-Boss/GPT-SoVITS/pull/1404) karışık metin hatalarını düzeltti.
+3. [PR 1355](https://github.com/RVC-Boss/GPT-SoVITS/pull/1356) WebUI'de ses işlenirken yolları otomatik olarak doldurdu.
+4. [Commit bce451a](https://github.com/RVC-Boss/GPT-SoVITS/commit/bce451a2d1641e581e200297d01f219aeaaf7299), [Commit 4c8b761](https://github.com/RVC-Boss/GPT-SoVITS/commit/4c8b7612206536b8b4435997acb69b25d93acb78) GPU tanıma mantığını optimize etti.
+5. [Commit 8a10147](https://github.com/RVC-Boss/GPT-SoVITS/commit/8a101474b5a4f913b4c94fca2e3ca87d0771bae3) Kantonca ASR desteği ekledi.
+6. GPT-SoVITS v2 desteği eklendi.
+7. [PR 1387](https://github.com/RVC-Boss/GPT-SoVITS/pull/1387) zamanlama mantığını optimize etti.
+
+### 20240821 Güncelleme
+
+1. [PR 1490](https://github.com/RVC-Boss/GPT-SoVITS/pull/1490) `fast_inference` dalını ana dala birleştir.
+2. [Issue 1508](https://github.com/RVC-Boss/GPT-SoVITS/issues/1508) SSML etiketlerini kullanarak sayıları, telefon numaralarını, tarihleri ve saatleri optimize etme desteği.
+3. [PR 1503](https://github.com/RVC-Boss/GPT-SoVITS/pull/1503) API düzeltildi ve optimize edildi.
+4. [PR 1422](https://github.com/RVC-Boss/GPT-SoVITS/pull/1422) Karıştırmak için yalnızca bir referans sesi yüklenebiliyordu hatası düzeltildi, çeşitli veri seti kontrolleri eklendi ve eksik dosyalar için uyarılar çıkar.
+
+### 20250211 Güncellemesi
+
+- [Wiki](https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)) GPT-SoVITS v3 Modeli Eklendi, SoVITS v3'ü ince ayar yapmak için 14GB GPU belleği gereklidir.
+
+### 20250212 Güncellemesi
+
+- [PR 2040](https://github.com/RVC-Boss/GPT-SoVITS/pull/2040) SoVITS v3'ü ince ayar yapmak için gradyan kontrol noktası ekledi, 12GB GPU belleği gereklidir.
+
+### 20250214 Güncellemesi
+
+- [PR 2047](https://github.com/RVC-Boss/GPT-SoVITS/pull/2047) Çok dilli karışık metin segmentasyon stratejisi **A**'yı optimize etti.
+  - `split-lang` bir dil segmentasyon aracı olarak eklendi ve çok dilli karışık metinlerin segmentasyon yeteneklerini iyileştirdi.
+
+### 20250217 Güncellemesi
+
+- [PR 2062](https://github.com/RVC-Boss/GPT-SoVITS/pull/2062) Metindeki sayılar ve İngilizceyi işleme mantığını optimize etti.
+
+### 20250218 Güncellemesi
+
+- [PR 2073](https://github.com/RVC-Boss/GPT-SoVITS/pull/2073) Çok dilli karışık metin segmentasyon stratejisi **B**'yi optimize etti.
+
+### 20250223 Güncellemesi
+
+1. SoVITS V3 için LoRA eğitimi, ince ayar yapmayı destekler. 8GB GPU belleği gereklidir ve sonuçlar tam parametreli ince ayar yapmaktan daha iyidir.
+2. [PR 2078](https://github.com/RVC-Boss/GPT-SoVITS/pull/2078) Mel Band RoFormer modelini vokal ve enstrüman ayrımı için ekledi.
+
+### 20250226 Güncellemesi
+
+1. [PR 2112](https://github.com/RVC-Boss/GPT-SoVITS/pull/2112) Windows'ta İngilizce olmayan dizinlerden kaynaklanan sorunları düzeltti.
+   - Korece için `langsegmenter` kullanımı ile ilgili sorun düzeltildi.
+2. [PR 2113](https://github.com/RVC-Boss/GPT-SoVITS/pull/2114) Windows'ta İngilizce olmayan dizinlerden kaynaklanan sorunları düzeltti.
+   - Korece/Japonca için `langsegmenter` kullanımı ile ilgili sorun düzeltildi.
+
+### 20250227 Güncellemesi
+
+- 24K sesli V3 modeliyle 24K ses oluştururken meydana gelen boğukluk sorununu hafifletmek için, 24K'dan 48K'ya ses süper çözünürlük modelleri eklendi. [Issue 2085](https://github.com/RVC-Boss/GPT-SoVITS/issues/2085), [Issue 2117](https://github.com/RVC-Boss/GPT-SoVITS/issues/2117) de bildirilen sorunlar.
\ No newline at end of file
diff --git a/docs/tr/README.md b/docs/tr/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..d59f66baf0da3e708c73d6e47d1113217508af48
--- /dev/null
+++ b/docs/tr/README.md
@@ -0,0 +1,385 @@
+<div align="center">
+
+<h1>GPT-SoVITS-WebUI</h1>
+Güçlü Birkaç Örnekli Ses Dönüştürme ve Metinden Konuşmaya Web Arayüzü.<br><br>
+
+[![madewithlove](https://img.shields.io/badge/made_with-%E2%9D%A4-red?style=for-the-badge&labelColor=orange)](https://github.com/RVC-Boss/GPT-SoVITS)
+
+<a href="https://trendshift.io/repositories/7033" target="_blank"><img src="https://trendshift.io/api/badge/repositories/7033" alt="RVC-Boss%2FGPT-SoVITS | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+
+<!-- img src="https://counter.seku.su/cmoe?name=gptsovits&theme=r34" /><br> -->
+
+[![Open In Colab](https://img.shields.io/badge/Colab-F9AB00?style=for-the-badge&logo=googlecolab&color=525252)](https://colab.research.google.com/github/RVC-Boss/GPT-SoVITS/blob/main/colab_webui.ipynb)
+[![License](https://img.shields.io/badge/LICENSE-MIT-green.svg?style=for-the-badge)](https://github.com/RVC-Boss/GPT-SoVITS/blob/main/LICENSE)
+[![Huggingface](https://img.shields.io/badge/🤗%20-online%20demo-yellow.svg?style=for-the-badge)](https://huggingface.co/spaces/lj1995/GPT-SoVITS-v2)
+[![Discord](https://img.shields.io/discord/1198701940511617164?color=%23738ADB&label=Discord&style=for-the-badge)](https://discord.gg/dnrgs5GHfG)
+
+[**English**](../../README.md) | [**中文简体**](../cn/README.md) | [**日本語**](../ja/README.md) | [**한국어**](../ko/README.md) | **Türkçe**
+
+</div>
+
+---
+
+## Özellikler:
+
+1. **Sıfır Örnekli Metinden Konuşmaya:** 5 saniyelik bir vokal örneği girin ve anında metinden konuşmaya dönüşümünü deneyimleyin.
+
+2. **Birkaç Örnekli Metinden Konuşmaya:** Daha iyi ses benzerliği ve gerçekçiliği için modeli yalnızca 1 dakikalık eğitim verisiyle ince ayarlayın.
+
+3. **Çapraz Dil Desteği:** Eğitim veri setinden farklı dillerde çıkarım, şu anda İngilizce, Japonca, Çince, Kantonca ve Koreceyi destekliyor.
+
+4. **Web Arayüzü Araçları:** Entegre araçlar arasında vokal eşliğinde ayırma, otomatik eğitim seti segmentasyonu, Çince ASR ve metin etiketleme bulunur ve yeni başlayanların eğitim veri setleri ve GPT/SoVITS modelleri oluşturmalarına yardımcı olur.
+
+**[Demo videomuzu](https://www.bilibili.com/video/BV12g4y1m7Uw) buradan izleyin!**
+
+Görünmeyen konuşmacılar birkaç örnekli ince ayar demosu:
+
+https://github.com/RVC-Boss/GPT-SoVITS/assets/129054828/05bee1fa-bdd8-4d85-9350-80c060ab47fb
+
+**Kullanıcı Kılavuzu: [简体中文](https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e) | [English](https://rentry.co/GPT-SoVITS-guide#/)**
+
+## Kurulum
+
+### Test Edilmiş Ortamlar
+
+| Python Version | PyTorch Version  | Device          |
+|----------------|------------------|-----------------|
+| Python 3.9     | PyTorch 2.0.1    | CUDA 11.8       |
+| Python 3.10.13 | PyTorch 2.1.2    | CUDA 12.3       |
+| Python 3.10.17 | PyTorch 2.5.1    | CUDA 12.4       |
+| Python 3.9     | PyTorch 2.5.1    | Apple silicon   |
+| Python 3.11    | PyTorch 2.6.0    | Apple silicon   |
+| Python 3.9     | PyTorch 2.2.2    | CPU             |
+
+### Windows
+
+Eğer bir Windows kullanıcısıysanız (win>=10 ile test edilmiştir), [entegre paketi indirin](https://huggingface.co/lj1995/GPT-SoVITS-windows-package/resolve/main/GPT-SoVITS-v3lora-20250228.7z?download=true) ve _go-webui.bat_ dosyasına çift tıklayarak GPT-SoVITS-WebUI'yi başlatın.
+
+### Linux
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### macOS
+
+**Not: Mac'lerde GPU'larla eğitilen modeller, diğer cihazlarda eğitilenlere göre önemli ölçüde daha düşük kalitede sonuç verir, bu nedenle geçici olarak CPU'lar kullanıyoruz.**
+
+1. `xcode-select --install` komutunu çalıştırarak Xcode komut satırı araçlarını yükleyin.
+2. Aşağıdaki komutları çalıştırarak programı yükleyin:
+
+```bash
+conda create -n GPTSoVits python=3.9
+conda activate GPTSoVits
+bash install.sh --source <HF|HF-Mirror|ModelScope> [--download-uvr5]
+```
+
+### El ile Yükleme
+
+#### FFmpeg'i Yükleme
+
+##### Conda Kullanıcıları
+
+```bash
+conda install ffmpeg
+```
+
+##### Ubuntu/Debian Kullanıcıları
+
+```bash
+sudo apt install ffmpeg
+sudo apt install libsox-dev
+conda install -c conda-forge 'ffmpeg<7'
+```
+
+##### Windows Kullanıcıları
+
+[ffmpeg.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffmpeg.exe) ve [ffprobe.exe](https://huggingface.co/lj1995/VoiceConversionWebUI/blob/main/ffprobe.exe) dosyalarını indirin ve GPT-SoVITS kök dizinine yerleştirin.
+
+##### MacOS Kullanıcıları
+
+```bash
+brew install ffmpeg
+```
+
+#### Bağımlılıkları Yükleme
+
+```bash
+pip install -r extra-req.txt --no-deps
+pip install -r requirements.txt
+```
+
+### Docker Kullanarak
+
+#### docker-compose.yaml yapılandırması
+
+0. Görüntü etiketleri hakkında: Kod tabanındaki hızlı güncellemeler ve görüntüleri paketleme ve test etme işleminin yavaş olması nedeniyle, lütfen şu anda paketlenmiş en son görüntüleri kontrol etmek için [Docker Hub](https://hub.docker.com/r/breakstring/gpt-sovits)(eski sürüm) adresini kontrol edin ve durumunuza göre seçim yapın veya alternatif olarak, kendi ihtiyaçlarınıza göre bir Dockerfile kullanarak yerel olarak oluşturun.
+1. Ortam Değişkenleri: 
+   - is_half: Yarım hassasiyet/çift hassasiyeti kontrol eder. Bu genellikle "SSL çıkarma" adımı sırasında 4-cnhubert/5-wav32k dizinleri altındaki içeriğin doğru şekilde oluşturulmamasının nedenidir. Gerçek durumunuza göre True veya False olarak ayarlayın.
+2. Birim Yapılandırması, Kapsayıcı içindeki uygulamanın kök dizini /workspace olarak ayarlanmıştır. Varsayılan docker-compose.yaml, içerik yükleme/indirme için bazı pratik örnekler listeler.
+3. shm_size: Windows üzerinde Docker Desktop için varsayılan kullanılabilir bellek çok küçüktür, bu da anormal işlemlere neden olabilir. Kendi durumunuza göre ayarlayın.
+4. Dağıtım bölümü altında, GPU ile ilgili ayarlar sisteminize ve gerçek koşullara göre dikkatlice ayarlanmalıdır.
+
+#### docker compose ile çalıştırma
+
+```
+docker compose -f "docker-compose.yaml" up -d
+```
+
+#### docker komutu ile çalıştırma
+
+Yukarıdaki gibi, ilgili parametreleri gerçek durumunuza göre değiştirin, ardından aşağıdaki komutu çalıştırın:
+
+```
+docker run --rm -it --gpus=all --env=is_half=False --volume=G:\GPT-SoVITS-DockerTest\output:/workspace/output --volume=G:\GPT-SoVITS-DockerTest\logs:/workspace/logs --volume=G:\GPT-SoVITS-DockerTest\SoVITS_weights:/workspace/SoVITS_weights --workdir=/workspace -p 9880:9880 -p 9871:9871 -p 9872:9872 -p 9873:9873 -p 9874:9874 --shm-size="16G" -d breakstring/gpt-sovits:xxxxx
+```
+
+## Önceden Eğitilmiş Modeller
+
+**Eğer `install.sh` başarıyla çalıştırılırsa, No.1,2,3 adımını atlayabilirsiniz.**
+
+1. [GPT-SoVITS Models](https://huggingface.co/lj1995/GPT-SoVITS) üzerinden önceden eğitilmiş modelleri indirip `GPT_SoVITS/pretrained_models` dizinine yerleştirin.
+
+2. [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) üzerinden modeli indirip sıkıştırmayı açın ve `G2PWModel` olarak yeniden adlandırın, ardından `GPT_SoVITS/text` dizinine yerleştirin. (Sadece Çince TTS için)
+
+3. UVR5 (Vokal/Enstrümantal Ayrımı & Yankı Giderme) için, [UVR5 Weights](https://huggingface.co/lj1995/VoiceConversionWebUI/tree/main/uvr5_weights) üzerinden modelleri indirip `tools/uvr5/uvr5_weights` dizinine yerleştirin.
+
+   - UVR5'te bs_roformer veya mel_band_roformer modellerini kullanıyorsanız, modeli ve ilgili yapılandırma dosyasını manuel olarak indirip `tools/UVR5/UVR5_weights` klasörüne yerleştirebilirsiniz. **Model dosyası ve yapılandırma dosyasının adı, uzantı dışında aynı olmalıdır**. Ayrıca, model ve yapılandırma dosyasının adlarında **"roformer"** kelimesi yer almalıdır, böylece roformer sınıfındaki bir model olarak tanınır.
+
+   - Model adı ve yapılandırma dosyası adı içinde **doğrudan model tipini belirtmek önerilir**. Örneğin: mel_mand_roformer, bs_roformer. Belirtilmezse, yapılandırma dosyasından özellikler karşılaştırılarak model tipi belirlenir. Örneğin, `bs_roformer_ep_368_sdr_12.9628.ckpt` modeli ve karşılık gelen yapılandırma dosyası `bs_roformer_ep_368_sdr_12.9628.yaml` bir çifttir. Aynı şekilde, `kim_mel_band_roformer.ckpt` ve `kim_mel_band_roformer.yaml` da bir çifttir.
+
+4. Çince ASR için, [Damo ASR Model](https://modelscope.cn/models/damo/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch/files), [Damo VAD Model](https://modelscope.cn/models/damo/speech_fsmn_vad_zh-cn-16k-common-pytorch/files) ve [Damo Punc Model](https://modelscope.cn/models/damo/punc_ct-transformer_zh-cn-common-vocab272727-pytorch/files) üzerinden modelleri indirip `tools/asr/models` dizinine yerleştirin.
+
+5. İngilizce veya Japonca ASR için, [Faster Whisper Large V3](https://huggingface.co/Systran/faster-whisper-large-v3) üzerinden modeli indirip `tools/asr/models` dizinine yerleştirin. Ayrıca, [diğer modeller](https://huggingface.co/Systran) benzer bir etki yaratabilir ve daha az disk alanı kaplayabilir.
+
+## Veri Seti Formatı
+
+TTS açıklama .list dosya formatı:
+
+```
+vocal_path|speaker_name|language|text
+```
+
+Dil sözlüğü:
+
+- 'zh': Çince
+- 'ja': Japonca
+- 'en': İngilizce
+- 'ko': Korece
+- 'yue': Kantonca
+
+Örnek:
+
+```
+D:\GPT-SoVITS\xxx/xxx.wav|xxx|en|I like playing Genshin.
+```
+
+## İnce Ayar ve Çıkarım
+
+### WebUI'yi Açın
+
+#### Entegre Paket Kullanıcıları
+
+`go-webui.bat` dosyasına çift tıklayın veya `go-webui.ps1` kullanın.
+V1'e geçmek istiyorsanız, `go-webui-v1.bat` dosyasına çift tıklayın veya `go-webui-v1.ps1` kullanın.
+
+#### Diğerleri
+
+```bash
+python webui.py <dil(isteğe bağlı)>
+```
+
+V1'e geçmek istiyorsanız,
+
+```bash
+python webui.py v1 <dil(isteğe bağlı)>
+```
+
+veya WebUI'de manuel olarak sürüm değiştirin.
+
+### İnce Ayar
+
+#### Yol Otomatik Doldurma artık destekleniyor
+
+    1. Ses yolunu doldurun
+    2. Sesi küçük parçalara ayırın
+    3. Gürültü azaltma (isteğe bağlı)
+    4. ASR
+    5. ASR transkripsiyonlarını düzeltin
+    6. Bir sonraki sekmeye geçin ve modeli ince ayar yapın
+
+### Çıkarım WebUI'sini Açın
+
+#### Entegre Paket Kullanıcıları
+
+`go-webui-v2.bat` dosyasına çift tıklayın veya `go-webui-v2.ps1` kullanın, ardından çıkarım webui'sini `1-GPT-SoVITS-TTS/1C-inference` adresinde açın.
+
+#### Diğerleri
+
+```bash
+python GPT_SoVITS/inference_webui.py <dil(isteğe bağlı)>
+```
+
+VEYA
+
+```bash
+python webui.py
+```
+
+ardından çıkarım webui'sini `1-GPT-SoVITS-TTS/1C-inference` adresinde açın.
+
+## V2 Sürüm Notları
+
+Yeni Özellikler:
+
+1. Korece ve Kantonca destekler
+
+2. Optimize edilmiş metin ön yüzü
+
+3. Önceden eğitilmiş model 2k saatten 5k saate kadar genişletildi
+
+4. Düşük kaliteli referans sesler için geliştirilmiş sentez kalitesi
+
+   [detaylar burada](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v2%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+V1 ortamından V2'yi kullanmak için:
+
+1. `pip install -r requirements.txt` ile bazı paketleri güncelleyin
+
+2. github'dan en son kodları klonlayın.
+
+3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main/gsv-v2final-pretrained) adresinden v2 önceden eğitilmiş modelleri indirin ve bunları `GPT_SoVITS\pretrained_models\gsv-v2final-pretrained` dizinine yerleştirin.
+
+   Ek olarak Çince V2: [G2PWModel.zip(HF)](https://huggingface.co/XXXXRT/GPT-SoVITS-Pretrained/resolve/main/G2PWModel.zip)| [G2PWModel.zip(ModelScope)](https://www.modelscope.cn/models/XXXXRT/GPT-SoVITS-Pretrained/resolve/master/G2PWModel.zip) (G2PW modellerini indirip, zipten çıkarıp, `G2PWModel` olarak yeniden adlandırıp `GPT_SoVITS/text` dizinine yerleştirin.)
+
+## V3 Sürüm Notları
+
+### Yeni Özellikler:
+
+1. **Tını benzerliği** daha yüksek olup, hedef konuşmacıyı yakınsamak için daha az eğitim verisi gerekmektedir (tını benzerliği, base model doğrudan kullanılacak şekilde fine-tuning yapılmadan önemli ölçüde iyileştirilmiştir).
+
+2. GPT modeli daha **kararlı** hale geldi, tekrarlar ve atlamalar azaldı ve **daha zengin duygusal ifadeler** ile konuşma üretmek daha kolay hale geldi.
+
+   [daha fazla detay](<https://github.com/RVC-Boss/GPT-SoVITS/wiki/GPT%E2%80%90SoVITS%E2%80%90v3%E2%80%90features-(%E6%96%B0%E7%89%B9%E6%80%A7)>)
+
+### v2 ortamında v3 kullanımı:
+
+1. `pip install -r requirements.txt` ile bazı paketleri güncelleyin.
+
+2. GitHub'dan en son kodları klonlayın.
+
+3. [huggingface](https://huggingface.co/lj1995/GPT-SoVITS/tree/main) üzerinden v3 önceden eğitilmiş modellerini (s1v3.ckpt, s2Gv3.pth ve models--nvidia--bigvgan_v2_24khz_100band_256x klasörünü) indirin ve `GPT_SoVITS\pretrained_models` dizinine yerleştirin.
+
+   ek: Ses Süper Çözünürlük modeli için [nasıl indirileceği](../../tools/AP_BWE_main/24kto48k/readme.txt) hakkında bilgi alabilirsiniz.
+
+## Yapılacaklar Listesi
+
+- [x] **Yüksek Öncelikli:**
+
+  - [x] Japonca ve İngilizceye yerelleştirme.
+  - [x] Kullanıcı kılavuzu.
+  - [x] Japonca ve İngilizce veri seti ince ayar eğitimi.
+
+- [ ] **Özellikler:**
+  - [x] Sıfır örnekli ses dönüştürme (5s) / birkaç örnekli ses dönüştürme (1dk).
+  - [x] Metinden konuşmaya konuşma hızı kontrolü.
+  - [ ] ~~Gelişmiş metinden konuşmaya duygu kontrolü.~~
+  - [ ] SoVITS token girdilerini kelime dağarcığı olasılık dağılımına değiştirme denemesi.
+  - [x] İngilizce ve Japonca metin ön ucunu iyileştirme.
+  - [ ] Küçük ve büyük boyutlu metinden konuşmaya modelleri geliştirme.
+  - [x] Colab betikleri.
+  - [ ] Eğitim veri setini genişletmeyi dene (2k saat -> 10k saat).
+  - [x] daha iyi sovits temel modeli (geliştirilmiş ses kalitesi)
+  - [ ] model karışımı
+
+## (Ekstra) Komut satırından çalıştırma yöntemi
+
+UVR5 için Web Arayüzünü açmak için komut satırını kullanın
+
+```
+python tools/uvr5/webui.py "<infer_device>" <is_half> <webui_port_uvr5>
+```
+
+<!-- Bir tarayıcı açamıyorsanız, UVR işleme için aşağıdaki formatı izleyin,Bu ses işleme için mdxnet kullanıyor
+```
+python mdxnet.py --model --input_root --output_vocal --output_ins --agg_level --format --device --is_half_precision
+``` -->
+
+Veri setinin ses segmentasyonu komut satırı kullanılarak bu şekilde yapılır
+
+```
+python audio_slicer.py \
+    --input_path "<orijinal_ses_dosyası_veya_dizininin_yolu>" \
+    --output_root "<alt_bölümlere_ayrılmış_ses_kliplerinin_kaydedileceği_dizin>" \
+    --threshold <ses_eşiği> \
+    --min_length <her_bir_alt_klibin_minimum_süresi> \
+    --min_interval <bitişik_alt_klipler_arasındaki_en_kısa_zaman_aralığı>
+    --hop_size <ses_eğrisini_hesaplamak_için_adım_boyutu>
+```
+
+Veri seti ASR işleme komut satırı kullanılarak bu şekilde yapılır (Yalnızca Çince)
+
+```
+python tools/asr/funasr_asr.py -i <girdi> -o <çıktı>
+```
+
+ASR işleme Faster_Whisper aracılığıyla gerçekleştirilir (Çince dışındaki ASR işaretleme)
+
+(İlerleme çubukları yok, GPU performansı zaman gecikmelerine neden olabilir)
+
+```
+python ./tools/asr/fasterwhisper_asr.py -i <girdi> -o <çıktı> -l <dil>
+```
+
+Özel bir liste kaydetme yolu etkinleştirildi
+
+## Katkı Verenler
+
+Özellikle aşağıdaki projelere ve katkıda bulunanlara teşekkür ederiz:
+
+### Teorik Araştırma
+
+- [ar-vits](https://github.com/innnky/ar-vits)
+- [SoundStorm](https://github.com/yangdongchao/SoundStorm/tree/master/soundstorm/s1/AR)
+- [vits](https://github.com/jaywalnut310/vits)
+- [TransferTTS](https://github.com/hcy71o/TransferTTS/blob/master/models.py#L556)
+- [contentvec](https://github.com/auspicious3000/contentvec/)
+- [hifi-gan](https://github.com/jik876/hifi-gan)
+- [fish-speech](https://github.com/fishaudio/fish-speech/blob/main/tools/llama/generate.py#L41)
+- [f5-TTS](https://github.com/SWivid/F5-TTS/blob/main/src/f5_tts/model/backbones/dit.py)
+- [shortcut flow matching](https://github.com/kvfrans/shortcut-models/blob/main/targets_shortcut.py)
+
+### Önceden Eğitilmiş Modeller
+
+- [Chinese Speech Pretrain](https://github.com/TencentGameMate/chinese_speech_pretrain)
+- [Chinese-Roberta-WWM-Ext-Large](https://huggingface.co/hfl/chinese-roberta-wwm-ext-large)
+- [BigVGAN](https://github.com/NVIDIA/BigVGAN)
+
+### Tahmin İçin Metin Ön Ucu
+
+- [paddlespeech zh_normalization](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/zh_normalization)
+- [split-lang](https://github.com/DoodleBears/split-lang)
+- [g2pW](https://github.com/GitYCC/g2pW)
+- [pypinyin-g2pW](https://github.com/mozillazg/pypinyin-g2pW)
+- [paddlespeech g2pw](https://github.com/PaddlePaddle/PaddleSpeech/tree/develop/paddlespeech/t2s/frontend/g2pw)
+
+### WebUI Araçları
+
+- [ultimatevocalremovergui](https://github.com/Anjok07/ultimatevocalremovergui)
+- [audio-slicer](https://github.com/openvpi/audio-slicer)
+- [SubFix](https://github.com/cronrpc/SubFix)
+- [FFmpeg](https://github.com/FFmpeg/FFmpeg)
+- [gradio](https://github.com/gradio-app/gradio)
+- [faster-whisper](https://github.com/SYSTRAN/faster-whisper)
+- [FunASR](https://github.com/alibaba-damo-academy/FunASR)
+- [AP-BWE](https://github.com/yxlu-0102/AP-BWE)
+
+@Naozumi520'ye Kantonca eğitim setini sağladığı ve Kantonca ile ilgili bilgiler konusunda rehberlik ettiği için minnettarım.
+
+## Tüm katkıda bulunanlara çabaları için teşekkürler
+
+<a href="https://github.com/RVC-Boss/GPT-SoVITS/graphs/contributors" target="_blank">
+  <img src="https://contrib.rocks/image?repo=RVC-Boss/GPT-SoVITS" />
+</a>
diff --git a/tools/AP_BWE_main/24kto48k/readme.txt b/tools/AP_BWE_main/24kto48k/readme.txt
new file mode 100644
index 0000000000000000000000000000000000000000..21cef4a287401fc9dca081d1e4f23d05c301ac88
--- /dev/null
+++ b/tools/AP_BWE_main/24kto48k/readme.txt
@@ -0,0 +1,11 @@
+For the inference of the v3 model, if you find that the generated audio sounds somewhat muffled, you can try using this audio super-resolution model.
+对于v3模型的推理，如果你发现生成的音频比较闷，可以尝试这个音频超分模型。
+
+put g_24kto48k.zip and config.json in this folder
+把g_24kto48k.zip and config.json下到这个文件夹
+
+download link 下载链接:
+https://drive.google.com/drive/folders/1IIYTf2zbJWzelu4IftKD6ooHloJ8mnZF?usp=share_link
+
+audio sr project page 音频超分项目主页:
+https://github.com/yxlu-0102/AP-BWE
diff --git a/tools/AP_BWE_main/LICENSE b/tools/AP_BWE_main/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..61a53b959de3005cda36b3bcbb46904fbffbcf9a
--- /dev/null
+++ b/tools/AP_BWE_main/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Ye-Xin Lu
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/tools/AP_BWE_main/README.md b/tools/AP_BWE_main/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..3ad1d4c38afe00ab12575a3fb2edfbd07706e382
--- /dev/null
+++ b/tools/AP_BWE_main/README.md
@@ -0,0 +1,91 @@
+# Towards High-Quality and Efficient Speech Bandwidth Extension with Parallel Amplitude and Phase Prediction
+### Ye-Xin Lu, Yang Ai, Hui-Peng Du, Zhen-Hua Ling
+
+**Abstract:** 
+Speech bandwidth extension (BWE) refers to widening the frequency bandwidth range of speech signals, enhancing the speech quality towards brighter and fuller.
+This paper proposes a generative adversarial network (GAN) based BWE model with parallel prediction of Amplitude and Phase spectra, named AP-BWE, which achieves both high-quality and efficient wideband speech waveform generation.
+The proposed AP-BWE generator is entirely based on convolutional neural networks (CNNs).
+It features a dual-stream architecture with mutual interaction, where the amplitude stream and the phase stream communicate with each other and respectively extend the high-frequency components from the input narrowband amplitude and phase spectra.
+To improve the naturalness of the extended speech signals, we employ a multi-period discriminator at the waveform level and design a pair of multi-resolution amplitude and phase discriminators at the spectral level, respectively.
+Experimental results demonstrate that our proposed AP-BWE achieves state-of-the-art performance in terms of speech quality for BWE tasks targeting sampling rates of both 16 kHz and 48 kHz. 
+In terms of generation efficiency, due to the all-convolutional architecture and all-frame-level operations, the proposed AP-BWE can generate 48 kHz waveform samples 292.3 times faster than real-time on a single RTX 4090 GPU and 18.1 times faster than real-time on a single CPU.
+Notably, to our knowledge, AP-BWE is the first to achieve the direct extension of the high-frequency phase spectrum, which is beneficial for improving the effectiveness of existing BWE methods.
+
+**We provide our implementation as open source in this repository. Audio samples can be found at the [demo website](http://yxlu-0102.github.io/AP-BWE).**
+
+
+## Pre-requisites
+0. Python >= 3.9.
+0. Clone this repository.
+0. Install python requirements. Please refer [requirements.txt](requirements.txt).
+0. Download datasets
+    1. Download and extract the [VCTK-0.92 dataset](https://datashare.ed.ac.uk/handle/10283/3443), and move its `wav48` directory into [VCTK-Corpus-0.92](VCTK-Corpus-0.92) and rename it as `wav48_origin`.
+    1. Trim the silence of the dataset, and the trimmed files will be saved to `wav48_silence_trimmed`.
+       ```
+       cd VCTK-Corpus-0.92
+       python flac2wav.py
+       ```
+    1. Move all the trimmed training files from `wav48_silence_trimmed` to [wav48/train](wav48/train) following the indexes in [training.txt](VCTK-Corpus-0.92/training.txt), and move all the untrimmed test files from  `wav48_origin` to [wav48/test](wav48/test) following the indexes in [test.txt](VCTK-Corpus-0.92/test.txt).
+
+## Training
+```
+cd train
+CUDA_VISIBLE_DEVICES=0 python train_16k.py --config [config file path]
+CUDA_VISIBLE_DEVICES=0 python train_48k.py --config [config file path]
+```
+Checkpoints and copies of the configuration file are saved in the `cp_model` directory by default.<br>
+You can change the path by using the `--checkpoint_path` option.
+Here is an example:
+```
+CUDA_VISIBLE_DEVICES=0 python train_16k.py --config ../configs/config_2kto16k.json --checkpoint_path ../checkpoints/AP-BWE_2kto16k
+```
+
+## Inference
+```
+cd inference
+python inference_16k.py --checkpoint_file [generator checkpoint file path]
+python inference_48k.py --checkpoint_file [generator checkpoint file path]
+```
+You can download the [pretrained weights](https://drive.google.com/drive/folders/1IIYTf2zbJWzelu4IftKD6ooHloJ8mnZF?usp=share_link) we provide and move all the files to the `checkpoints` directory.
+<br>
+Generated wav files are saved in `generated_files` by default.
+You can change the path by adding `--output_dir` option.
+Here is an example:
+```
+python inference_16k.py --checkpoint_file ../checkpoints/2kto16k/g_2kto16k --output_dir ../generated_files/2kto16k
+```
+
+## Model Structure
+![model](Figures/model.png)
+
+## Comparison with other speech BWE methods
+### 2k/4k/8kHz to 16kHz
+<p align="center">
+<img src="Figures/table_16k.png" alt="comparison" width="90%"/>
+</p>
+
+### 8k/12k/16/24kHz to 16kHz
+<p align="center">
+<img src="Figures/table_48k.png" alt="comparison" width="100%"/>
+</p>
+
+## Acknowledgements
+We referred to [HiFi-GAN](https://github.com/jik876/hifi-gan) and [NSPP](https://github.com/YangAi520/NSPP) to implement this.
+
+## Citation
+```
+@article{lu2024towards,
+  title={Towards high-quality and efficient speech bandwidth extension with parallel amplitude and phase prediction},
+  author={Lu, Ye-Xin and Ai, Yang and Du, Hui-Peng and Ling, Zhen-Hua},
+  journal={arXiv preprint arXiv:2401.06387},
+  year={2024}
+}
+
+@inproceedings{lu2024multi,
+  title={Multi-Stage Speech Bandwidth Extension with Flexible Sampling Rate Control},
+  author={Lu, Ye-Xin and Ai, Yang and Sheng, Zheng-Yan and Ling, Zhen-Hua},
+  booktitle={Proc. Interspeech},
+  pages={2270--2274},
+  year={2024}
+}
+```
diff --git a/tools/AP_BWE_main/datasets1/__init__.py b/tools/AP_BWE_main/datasets1/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc
--- /dev/null
+++ b/tools/AP_BWE_main/datasets1/__init__.py
@@ -0,0 +1 @@
+
diff --git a/tools/AP_BWE_main/datasets1/dataset.py b/tools/AP_BWE_main/datasets1/dataset.py
new file mode 100644
index 0000000000000000000000000000000000000000..40f993b55808f7b118559e62ce0627cc0eead6e1
--- /dev/null
+++ b/tools/AP_BWE_main/datasets1/dataset.py
@@ -0,0 +1,108 @@
+import os
+import random
+import torch
+import torchaudio
+import torch.utils.data
+import torchaudio.functional as aF
+
+
+def amp_pha_stft(audio, n_fft, hop_size, win_size, center=True):
+    hann_window = torch.hann_window(win_size).to(audio.device)
+    stft_spec = torch.stft(
+        audio,
+        n_fft,
+        hop_length=hop_size,
+        win_length=win_size,
+        window=hann_window,
+        center=center,
+        pad_mode="reflect",
+        normalized=False,
+        return_complex=True,
+    )
+    log_amp = torch.log(torch.abs(stft_spec) + 1e-4)
+    pha = torch.angle(stft_spec)
+
+    com = torch.stack((torch.exp(log_amp) * torch.cos(pha), torch.exp(log_amp) * torch.sin(pha)), dim=-1)
+
+    return log_amp, pha, com
+
+
+def amp_pha_istft(log_amp, pha, n_fft, hop_size, win_size, center=True):
+    amp = torch.exp(log_amp)
+    com = torch.complex(amp * torch.cos(pha), amp * torch.sin(pha))
+    hann_window = torch.hann_window(win_size).to(com.device)
+    audio = torch.istft(com, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window, center=center)
+
+    return audio
+
+
+def get_dataset_filelist(a):
+    with open(a.input_training_file, "r", encoding="utf-8") as fi:
+        training_indexes = [x.split("|")[0] for x in fi.read().split("\n") if len(x) > 0]
+
+    with open(a.input_validation_file, "r", encoding="utf-8") as fi:
+        validation_indexes = [x.split("|")[0] for x in fi.read().split("\n") if len(x) > 0]
+
+    return training_indexes, validation_indexes
+
+
+class Dataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        training_indexes,
+        wavs_dir,
+        segment_size,
+        hr_sampling_rate,
+        lr_sampling_rate,
+        split=True,
+        shuffle=True,
+        n_cache_reuse=1,
+        device=None,
+    ):
+        self.audio_indexes = training_indexes
+        random.seed(1234)
+        if shuffle:
+            random.shuffle(self.audio_indexes)
+        self.wavs_dir = wavs_dir
+        self.segment_size = segment_size
+        self.hr_sampling_rate = hr_sampling_rate
+        self.lr_sampling_rate = lr_sampling_rate
+        self.split = split
+        self.cached_wav = None
+        self.n_cache_reuse = n_cache_reuse
+        self._cache_ref_count = 0
+        self.device = device
+
+    def __getitem__(self, index):
+        filename = self.audio_indexes[index]
+        if self._cache_ref_count == 0:
+            audio, orig_sampling_rate = torchaudio.load(os.path.join(self.wavs_dir, filename + ".wav"))
+            self.cached_wav = audio
+            self._cache_ref_count = self.n_cache_reuse
+        else:
+            audio = self.cached_wav
+            self._cache_ref_count -= 1
+
+        if orig_sampling_rate == self.hr_sampling_rate:
+            audio_hr = audio
+        else:
+            audio_hr = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.hr_sampling_rate)
+
+        audio_lr = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.lr_sampling_rate)
+        audio_lr = aF.resample(audio_lr, orig_freq=self.lr_sampling_rate, new_freq=self.hr_sampling_rate)
+        audio_lr = audio_lr[:, : audio_hr.size(1)]
+
+        if self.split:
+            if audio_hr.size(1) >= self.segment_size:
+                max_audio_start = audio_hr.size(1) - self.segment_size
+                audio_start = random.randint(0, max_audio_start)
+                audio_hr = audio_hr[:, audio_start : audio_start + self.segment_size]
+                audio_lr = audio_lr[:, audio_start : audio_start + self.segment_size]
+            else:
+                audio_hr = torch.nn.functional.pad(audio_hr, (0, self.segment_size - audio_hr.size(1)), "constant")
+                audio_lr = torch.nn.functional.pad(audio_lr, (0, self.segment_size - audio_lr.size(1)), "constant")
+
+        return (audio_hr.squeeze(), audio_lr.squeeze())
+
+    def __len__(self):
+        return len(self.audio_indexes)
diff --git a/tools/AP_BWE_main/models/__init__.py b/tools/AP_BWE_main/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8b137891791fe96927ad78e64b0aad7bded08bdc
--- /dev/null
+++ b/tools/AP_BWE_main/models/__init__.py
@@ -0,0 +1 @@
+
diff --git a/tools/AP_BWE_main/models/model.py b/tools/AP_BWE_main/models/model.py
new file mode 100644
index 0000000000000000000000000000000000000000..e538600162ba26f26978dcee660e3671c0249ee9
--- /dev/null
+++ b/tools/AP_BWE_main/models/model.py
@@ -0,0 +1,464 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch.nn.utils import weight_norm, spectral_norm
+
+
+# from utils import init_weights, get_padding
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+        m.weight.data.normal_(mean, std)
+
+
+import numpy as np
+from typing import Tuple, List
+
+LRELU_SLOPE = 0.1
+
+
+class ConvNeXtBlock(nn.Module):
+    """ConvNeXt Block adapted from https://github.com/facebookresearch/ConvNeXt to 1D audio signal.
+
+    Args:
+        dim (int): Number of input channels.
+        intermediate_dim (int): Dimensionality of the intermediate layer.
+        layer_scale_init_value (float, optional): Initial value for the layer scale. None means no scaling.
+            Defaults to None.
+        adanorm_num_embeddings (int, optional): Number of embeddings for AdaLayerNorm.
+            None means non-conditional LayerNorm. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        layer_scale_init_value=None,
+        adanorm_num_embeddings=None,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv1d(dim, dim, kernel_size=7, padding=3, groups=dim)  # depthwise conv
+        self.adanorm = adanorm_num_embeddings is not None
+
+        self.norm = nn.LayerNorm(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(dim, dim * 3)  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(dim * 3, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones(dim), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+
+    def forward(self, x, cond_embedding_id=None):
+        residual = x
+        x = self.dwconv(x)
+        x = x.transpose(1, 2)  # (B, C, T) -> (B, T, C)
+        if self.adanorm:
+            assert cond_embedding_id is not None
+            x = self.norm(x, cond_embedding_id)
+        else:
+            x = self.norm(x)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.transpose(1, 2)  # (B, T, C) -> (B, C, T)
+
+        x = residual + x
+        return x
+
+
+class APNet_BWE_Model(torch.nn.Module):
+    def __init__(self, h):
+        super(APNet_BWE_Model, self).__init__()
+        self.h = h
+        self.adanorm_num_embeddings = None
+        layer_scale_init_value = 1 / h.ConvNeXt_layers
+
+        self.conv_pre_mag = nn.Conv1d(h.n_fft // 2 + 1, h.ConvNeXt_channels, 7, 1, padding=get_padding(7, 1))
+        self.norm_pre_mag = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
+        self.conv_pre_pha = nn.Conv1d(h.n_fft // 2 + 1, h.ConvNeXt_channels, 7, 1, padding=get_padding(7, 1))
+        self.norm_pre_pha = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
+
+        self.convnext_mag = nn.ModuleList(
+            [
+                ConvNeXtBlock(
+                    dim=h.ConvNeXt_channels,
+                    layer_scale_init_value=layer_scale_init_value,
+                    adanorm_num_embeddings=self.adanorm_num_embeddings,
+                )
+                for _ in range(h.ConvNeXt_layers)
+            ]
+        )
+
+        self.convnext_pha = nn.ModuleList(
+            [
+                ConvNeXtBlock(
+                    dim=h.ConvNeXt_channels,
+                    layer_scale_init_value=layer_scale_init_value,
+                    adanorm_num_embeddings=self.adanorm_num_embeddings,
+                )
+                for _ in range(h.ConvNeXt_layers)
+            ]
+        )
+
+        self.norm_post_mag = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
+        self.norm_post_pha = nn.LayerNorm(h.ConvNeXt_channels, eps=1e-6)
+        self.apply(self._init_weights)
+        self.linear_post_mag = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
+        self.linear_post_pha_r = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
+        self.linear_post_pha_i = nn.Linear(h.ConvNeXt_channels, h.n_fft // 2 + 1)
+
+    def _init_weights(self, m):
+        if isinstance(m, (nn.Conv1d, nn.Linear)):
+            nn.init.trunc_normal_(m.weight, std=0.02)
+            nn.init.constant_(m.bias, 0)
+
+    def forward(self, mag_nb, pha_nb):
+        x_mag = self.conv_pre_mag(mag_nb)
+        x_pha = self.conv_pre_pha(pha_nb)
+        x_mag = self.norm_pre_mag(x_mag.transpose(1, 2)).transpose(1, 2)
+        x_pha = self.norm_pre_pha(x_pha.transpose(1, 2)).transpose(1, 2)
+
+        for conv_block_mag, conv_block_pha in zip(self.convnext_mag, self.convnext_pha):
+            x_mag = x_mag + x_pha
+            x_pha = x_pha + x_mag
+            x_mag = conv_block_mag(x_mag, cond_embedding_id=None)
+            x_pha = conv_block_pha(x_pha, cond_embedding_id=None)
+
+        x_mag = self.norm_post_mag(x_mag.transpose(1, 2))
+        mag_wb = mag_nb + self.linear_post_mag(x_mag).transpose(1, 2)
+
+        x_pha = self.norm_post_pha(x_pha.transpose(1, 2))
+        x_pha_r = self.linear_post_pha_r(x_pha)
+        x_pha_i = self.linear_post_pha_i(x_pha)
+        pha_wb = torch.atan2(x_pha_i, x_pha_r).transpose(1, 2)
+
+        com_wb = torch.stack((torch.exp(mag_wb) * torch.cos(pha_wb), torch.exp(mag_wb) * torch.sin(pha_wb)), dim=-1)
+
+        return mag_wb, pha_wb, com_wb
+
+
+class DiscriminatorP(torch.nn.Module):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(nn.Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+                norm_f(nn.Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+                norm_f(nn.Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+                norm_f(nn.Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(5, 1), 0))),
+                norm_f(nn.Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
+            ]
+        )
+        self.conv_post = norm_f(nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for i, l in enumerate(self.convs):
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            if i > 0:
+                fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(torch.nn.Module):
+    def __init__(self):
+        super(MultiPeriodDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorP(2),
+                DiscriminatorP(3),
+                DiscriminatorP(5),
+                DiscriminatorP(7),
+                DiscriminatorP(11),
+            ]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class MultiResolutionAmplitudeDiscriminator(nn.Module):
+    def __init__(
+        self,
+        resolutions: Tuple[Tuple[int, int, int]] = ((512, 128, 512), (1024, 256, 1024), (2048, 512, 2048)),
+        num_embeddings: int = None,
+    ):
+        super().__init__()
+        self.discriminators = nn.ModuleList(
+            [DiscriminatorAR(resolution=r, num_embeddings=num_embeddings) for r in resolutions]
+        )
+
+    def forward(
+        self, y: torch.Tensor, y_hat: torch.Tensor, bandwidth_id: torch.Tensor = None
+    ) -> 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(x=y, cond_embedding_id=bandwidth_id)
+            y_d_g, fmap_g = d(x=y_hat, cond_embedding_id=bandwidth_id)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorAR(nn.Module):
+    def __init__(
+        self,
+        resolution: Tuple[int, int, int],
+        channels: int = 64,
+        in_channels: int = 1,
+        num_embeddings: int = None,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(nn.Conv2d(in_channels, channels, kernel_size=(7, 5), stride=(2, 2), padding=(3, 2))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 1), padding=(2, 1))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 2), padding=(2, 1))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 1), padding=1)),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 2), padding=1)),
+            ]
+        )
+        if num_embeddings is not None:
+            self.emb = torch.nn.Embedding(num_embeddings=num_embeddings, embedding_dim=channels)
+            torch.nn.init.zeros_(self.emb.weight)
+        self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), padding=(1, 1)))
+
+    def forward(
+        self, x: torch.Tensor, cond_embedding_id: torch.Tensor = None
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        fmap = []
+        x = x.squeeze(1)
+
+        x = self.spectrogram(x)
+        x = x.unsqueeze(1)
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        if cond_embedding_id is not None:
+            emb = self.emb(cond_embedding_id)
+            h = (emb.view(1, -1, 1, 1) * x).sum(dim=1, keepdims=True)
+        else:
+            h = 0
+        x = self.conv_post(x)
+        fmap.append(x)
+        x += h
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+    def spectrogram(self, x: torch.Tensor) -> torch.Tensor:
+        n_fft, hop_length, win_length = self.resolution
+        amplitude_spectrogram = torch.stft(
+            x,
+            n_fft=n_fft,
+            hop_length=hop_length,
+            win_length=win_length,
+            window=None,  # interestingly rectangular window kind of works here
+            center=True,
+            return_complex=True,
+        ).abs()
+
+        return amplitude_spectrogram
+
+
+class MultiResolutionPhaseDiscriminator(nn.Module):
+    def __init__(
+        self,
+        resolutions: Tuple[Tuple[int, int, int]] = ((512, 128, 512), (1024, 256, 1024), (2048, 512, 2048)),
+        num_embeddings: int = None,
+    ):
+        super().__init__()
+        self.discriminators = nn.ModuleList(
+            [DiscriminatorPR(resolution=r, num_embeddings=num_embeddings) for r in resolutions]
+        )
+
+    def forward(
+        self, y: torch.Tensor, y_hat: torch.Tensor, bandwidth_id: torch.Tensor = None
+    ) -> 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(x=y, cond_embedding_id=bandwidth_id)
+            y_d_g, fmap_g = d(x=y_hat, cond_embedding_id=bandwidth_id)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorPR(nn.Module):
+    def __init__(
+        self,
+        resolution: Tuple[int, int, int],
+        channels: int = 64,
+        in_channels: int = 1,
+        num_embeddings: int = None,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(nn.Conv2d(in_channels, channels, kernel_size=(7, 5), stride=(2, 2), padding=(3, 2))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 1), padding=(2, 1))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=(5, 3), stride=(2, 2), padding=(2, 1))),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 1), padding=1)),
+                weight_norm(nn.Conv2d(channels, channels, kernel_size=3, stride=(2, 2), padding=1)),
+            ]
+        )
+        if num_embeddings is not None:
+            self.emb = torch.nn.Embedding(num_embeddings=num_embeddings, embedding_dim=channels)
+            torch.nn.init.zeros_(self.emb.weight)
+        self.conv_post = weight_norm(nn.Conv2d(channels, 1, (3, 3), padding=(1, 1)))
+
+    def forward(
+        self, x: torch.Tensor, cond_embedding_id: torch.Tensor = None
+    ) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        fmap = []
+        x = x.squeeze(1)
+
+        x = self.spectrogram(x)
+        x = x.unsqueeze(1)
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, LRELU_SLOPE)
+            fmap.append(x)
+        if cond_embedding_id is not None:
+            emb = self.emb(cond_embedding_id)
+            h = (emb.view(1, -1, 1, 1) * x).sum(dim=1, keepdims=True)
+        else:
+            h = 0
+        x = self.conv_post(x)
+        fmap.append(x)
+        x += h
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+    def spectrogram(self, x: torch.Tensor) -> torch.Tensor:
+        n_fft, hop_length, win_length = self.resolution
+        phase_spectrogram = torch.stft(
+            x,
+            n_fft=n_fft,
+            hop_length=hop_length,
+            win_length=win_length,
+            window=None,  # interestingly rectangular window kind of works here
+            center=True,
+            return_complex=True,
+        ).angle()
+
+        return phase_spectrogram
+
+
+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):
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss
+
+
+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):
+        r_loss = torch.mean(torch.clamp(1 - dr, min=0))
+        g_loss = torch.mean(torch.clamp(1 + dg, min=0))
+        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:
+        l = torch.mean(torch.clamp(1 - dg, min=0))
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses
+
+
+def phase_losses(phase_r, phase_g):
+    ip_loss = torch.mean(anti_wrapping_function(phase_r - phase_g))
+    gd_loss = torch.mean(anti_wrapping_function(torch.diff(phase_r, dim=1) - torch.diff(phase_g, dim=1)))
+    iaf_loss = torch.mean(anti_wrapping_function(torch.diff(phase_r, dim=2) - torch.diff(phase_g, dim=2)))
+
+    return ip_loss, gd_loss, iaf_loss
+
+
+def anti_wrapping_function(x):
+    return torch.abs(x - torch.round(x / (2 * np.pi)) * 2 * np.pi)
+
+
+def stft_mag(audio, n_fft=2048, hop_length=512):
+    hann_window = torch.hann_window(n_fft).to(audio.device)
+    stft_spec = torch.stft(audio, n_fft, hop_length, window=hann_window, return_complex=True)
+    stft_mag = torch.abs(stft_spec)
+    return stft_mag
+
+
+def cal_snr(pred, target):
+    snr = (20 * torch.log10(torch.norm(target, dim=-1) / torch.norm(pred - target, dim=-1).clamp(min=1e-8))).mean()
+    return snr
+
+
+def cal_lsd(pred, target):
+    sp = torch.log10(stft_mag(pred).square().clamp(1e-8))
+    st = torch.log10(stft_mag(target).square().clamp(1e-8))
+    return (sp - st).square().mean(dim=1).sqrt().mean()
diff --git a/tools/__init__.py b/tools/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/tools/asr/config.py b/tools/asr/config.py
new file mode 100644
index 0000000000000000000000000000000000000000..c04069b2c58bbc1bf00572f995623223131f897a
--- /dev/null
+++ b/tools/asr/config.py
@@ -0,0 +1,36 @@
+import os
+
+
+def check_fw_local_models():
+    """
+    启动时检查本地是否有 Faster Whisper 模型.
+    """
+    model_size_list = [
+        "tiny",
+        "tiny.en",
+        "base",
+        "base.en",
+        "small",
+        "small.en",
+        "medium",
+        "medium.en",
+        "large",
+        "large-v1",
+        "large-v2",
+        "large-v3",
+    ]
+    for i, size in enumerate(model_size_list):
+        if os.path.exists(f"tools/asr/models/faster-whisper-{size}"):
+            model_size_list[i] = size + "-local"
+    return model_size_list
+
+
+asr_dict = {
+    "达摩 ASR (中文)": {"lang": ["zh", "yue"], "size": ["large"], "path": "funasr_asr.py", "precision": ["float32"]},
+    "Faster Whisper (多语种)": {
+        "lang": ["auto", "zh", "en", "ja", "ko", "yue"],
+        "size": check_fw_local_models(),
+        "path": "fasterwhisper_asr.py",
+        "precision": ["float32", "float16", "int8"],
+    },
+}
diff --git a/tools/asr/fasterwhisper_asr.py b/tools/asr/fasterwhisper_asr.py
new file mode 100644
index 0000000000000000000000000000000000000000..e570f174805922d702592af83e37b41c59f4e0b9
--- /dev/null
+++ b/tools/asr/fasterwhisper_asr.py
@@ -0,0 +1,129 @@
+import argparse
+import os
+import traceback
+
+os.environ["HF_ENDPOINT"] = "https://hf-mirror.com"
+os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
+
+import torch
+from faster_whisper import WhisperModel
+from tqdm import tqdm
+
+from tools.asr.config import check_fw_local_models
+
+# fmt: off
+language_code_list = [
+    "af", "am", "ar", "as", "az", 
+    "ba", "be", "bg", "bn", "bo", 
+    "br", "bs", "ca", "cs", "cy", 
+    "da", "de", "el", "en", "es", 
+    "et", "eu", "fa", "fi", "fo", 
+    "fr", "gl", "gu", "ha", "haw", 
+    "he", "hi", "hr", "ht", "hu", 
+    "hy", "id", "is", "it", "ja", 
+    "jw", "ka", "kk", "km", "kn", 
+    "ko", "la", "lb", "ln", "lo", 
+    "lt", "lv", "mg", "mi", "mk", 
+    "ml", "mn", "mr", "ms", "mt", 
+    "my", "ne", "nl", "nn", "no", 
+    "oc", "pa", "pl", "ps", "pt", 
+    "ro", "ru", "sa", "sd", "si", 
+    "sk", "sl", "sn", "so", "sq", 
+    "sr", "su", "sv", "sw", "ta", 
+    "te", "tg", "th", "tk", "tl", 
+    "tr", "tt", "uk", "ur", "uz", 
+    "vi", "yi", "yo", "zh", "yue",
+    "auto"] 
+# fmt: on
+
+
+def execute_asr(input_folder, output_folder, model_size, language, precision):
+    if "-local" in model_size:
+        model_size = model_size[:-6]
+        model_path = f"tools/asr/models/faster-whisper-{model_size}"
+    else:
+        model_path = model_size
+    if language == "auto":
+        language = None  # 不设置语种由模型自动输出概率最高的语种
+    print("loading faster whisper model:", model_size, model_path)
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    try:
+        model = WhisperModel(model_path, device=device, compute_type=precision)
+    except:
+        return print(traceback.format_exc())
+
+    input_file_names = os.listdir(input_folder)
+    input_file_names.sort()
+
+    output = []
+    output_file_name = os.path.basename(input_folder)
+
+    for file_name in tqdm(input_file_names):
+        try:
+            file_path = os.path.join(input_folder, file_name)
+            segments, info = model.transcribe(
+                audio=file_path,
+                beam_size=5,
+                vad_filter=True,
+                vad_parameters=dict(min_silence_duration_ms=700),
+                language=language,
+            )
+            text = ""
+
+            if info.language == "zh":
+                print("检测为中文文本, 转 FunASR 处理")
+                if "only_asr" not in globals():
+                    from tools.asr.funasr_asr import only_asr  # 如果用英文就不需要导入下载模型
+                text = only_asr(file_path, language=info.language.lower())
+
+            if text == "":
+                for segment in segments:
+                    text += segment.text
+            output.append(f"{file_path}|{output_file_name}|{info.language.upper()}|{text}")
+        except:
+            print(traceback.format_exc())
+
+    output_folder = output_folder or "output/asr_opt"
+    os.makedirs(output_folder, exist_ok=True)
+    output_file_path = os.path.abspath(f"{output_folder}/{output_file_name}.list")
+
+    with open(output_file_path, "w", encoding="utf-8") as f:
+        f.write("\n".join(output))
+        print(f"ASR 任务完成->标注文件路径: {output_file_path}\n")
+    return output_file_path
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-i", "--input_folder", type=str, required=True, help="Path to the folder containing WAV files."
+    )
+    parser.add_argument("-o", "--output_folder", type=str, required=True, help="Output folder to store transcriptions.")
+    parser.add_argument(
+        "-s",
+        "--model_size",
+        type=str,
+        default="large-v3",
+        choices=check_fw_local_models(),
+        help="Model Size of Faster Whisper",
+    )
+    parser.add_argument(
+        "-l", "--language", type=str, default="ja", choices=language_code_list, help="Language of the audio files."
+    )
+    parser.add_argument(
+        "-p",
+        "--precision",
+        type=str,
+        default="float16",
+        choices=["float16", "float32", "int8"],
+        help="fp16, int8 or fp32",
+    )
+
+    cmd = parser.parse_args()
+    output_file_path = execute_asr(
+        input_folder=cmd.input_folder,
+        output_folder=cmd.output_folder,
+        model_size=cmd.model_size,
+        language=cmd.language,
+        precision=cmd.precision,
+    )
diff --git a/tools/asr/funasr_asr.py b/tools/asr/funasr_asr.py
new file mode 100644
index 0000000000000000000000000000000000000000..b0ffceb046ab9515d2f717e7450bef6b7799fe3d
--- /dev/null
+++ b/tools/asr/funasr_asr.py
@@ -0,0 +1,118 @@
+# -*- coding:utf-8 -*-
+
+import argparse
+import os
+import traceback
+
+# from funasr.utils import version_checker
+# version_checker.check_for_update = lambda: None
+from funasr import AutoModel
+from tqdm import tqdm
+
+funasr_models = {}  # 存储模型避免重复加载
+
+
+def only_asr(input_file, language):
+    try:
+        model = create_model(language)
+        text = model.generate(input=input_file)[0]["text"]
+    except:
+        text = ""
+        print(traceback.format_exc())
+    return text
+
+
+def create_model(language="zh"):
+    path_vad = "tools/asr/models/speech_fsmn_vad_zh-cn-16k-common-pytorch"
+    path_punc = "tools/asr/models/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
+    path_vad = path_vad if os.path.exists(path_vad) else "iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"
+    path_punc = path_punc if os.path.exists(path_punc) else "iic/punc_ct-transformer_zh-cn-common-vocab272727-pytorch"
+    vad_model_revision = punc_model_revision = "v2.0.4"
+
+    if language == "zh":
+        path_asr = "tools/asr/models/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+        path_asr = (
+            path_asr
+            if os.path.exists(path_asr)
+            else "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
+        )
+        model_revision = "v2.0.4"
+    elif language == "yue":
+        path_asr = "tools/asr/models/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online"
+        path_asr = (
+            path_asr
+            if os.path.exists(path_asr)
+            else "iic/speech_UniASR_asr_2pass-cantonese-CHS-16k-common-vocab1468-tensorflow1-online"
+        )
+        model_revision = "master"
+        path_vad = path_punc = None
+        vad_model_revision = punc_model_revision = None
+        ###友情提示：粤语带VAD识别可能会有少量shape不对报错的，但是不带VAD可以.不带vad只能分阶段单独加标点。不过标点模型对粤语效果真的不行…
+    else:
+        raise ValueError("FunASR 不支持该语言" + ": " + language)
+
+    if language in funasr_models:
+        return funasr_models[language]
+    else:
+        model = AutoModel(
+            model=path_asr,
+            model_revision=model_revision,
+            vad_model=path_vad,
+            vad_model_revision=vad_model_revision,
+            punc_model=path_punc,
+            punc_model_revision=punc_model_revision,
+        )
+        print(f"FunASR 模型加载完成: {language.upper()}")
+
+        funasr_models[language] = model
+        return model
+
+
+def execute_asr(input_folder, output_folder, model_size, language):
+    input_file_names = os.listdir(input_folder)
+    input_file_names.sort()
+
+    output = []
+    output_file_name = os.path.basename(input_folder)
+
+    model = create_model(language)
+
+    for file_name in tqdm(input_file_names):
+        try:
+            print("\n" + file_name)
+            file_path = os.path.join(input_folder, file_name)
+            text = model.generate(input=file_path)[0]["text"]
+            output.append(f"{file_path}|{output_file_name}|{language.upper()}|{text}")
+        except:
+            print(traceback.format_exc())
+
+    output_folder = output_folder or "output/asr_opt"
+    os.makedirs(output_folder, exist_ok=True)
+    output_file_path = os.path.abspath(f"{output_folder}/{output_file_name}.list")
+
+    with open(output_file_path, "w", encoding="utf-8") as f:
+        f.write("\n".join(output))
+        print(f"ASR 任务完成->标注文件路径: {output_file_path}\n")
+    return output_file_path
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-i", "--input_folder", type=str, required=True, help="Path to the folder containing WAV files."
+    )
+    parser.add_argument("-o", "--output_folder", type=str, required=True, help="Output folder to store transcriptions.")
+    parser.add_argument("-s", "--model_size", type=str, default="large", help="Model Size of FunASR is Large")
+    parser.add_argument(
+        "-l", "--language", type=str, default="zh", choices=["zh", "yue", "auto"], help="Language of the audio files."
+    )
+    parser.add_argument(
+        "-p", "--precision", type=str, default="float16", choices=["float16", "float32"], help="fp16 or fp32"
+    )  # 还没接入
+    cmd = parser.parse_args()
+    execute_asr(
+        input_folder=cmd.input_folder,
+        output_folder=cmd.output_folder,
+        model_size=cmd.model_size,
+        language=cmd.language,
+    )
diff --git a/tools/asr/models/.gitignore b/tools/asr/models/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..c96a04f008ee21e260b28f7701595ed59e2839e3
--- /dev/null
+++ b/tools/asr/models/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
\ No newline at end of file
diff --git a/tools/audio_sr.py b/tools/audio_sr.py
new file mode 100644
index 0000000000000000000000000000000000000000..58df6d201d2a044c246bc7fe69559f508cd95dfb
--- /dev/null
+++ b/tools/audio_sr.py
@@ -0,0 +1,50 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+import sys
+import os
+
+AP_BWE_main_dir_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "AP_BWE_main")
+sys.path.append(AP_BWE_main_dir_path)
+import json
+import torch
+import torchaudio.functional as aF
+# from attrdict import AttrDict####will be bug in py3.10
+
+from datasets1.dataset import amp_pha_stft, amp_pha_istft
+from models.model import APNet_BWE_Model
+
+
+class AP_BWE:
+    def __init__(self, device, DictToAttrRecursive, checkpoint_file=None):
+        if checkpoint_file == None:
+            checkpoint_file = "%s/24kto48k/g_24kto48k.zip" % (AP_BWE_main_dir_path)
+            if os.path.exists(checkpoint_file) == False:
+                raise FileNotFoundError
+        config_file = os.path.join(os.path.split(checkpoint_file)[0], "config.json")
+        with open(config_file) as f:
+            data = f.read()
+        json_config = json.loads(data)
+        # h = AttrDict(json_config)
+        h = DictToAttrRecursive(json_config)
+        model = APNet_BWE_Model(h).to(device)
+        state_dict = torch.load(checkpoint_file, map_location="cpu", weights_only=False)
+        model.load_state_dict(state_dict["generator"])
+        model.eval()
+        self.device = device
+        self.model = model
+        self.h = h
+
+    def to(self, *arg, **kwargs):
+        self.model.to(*arg, **kwargs)
+        self.device = self.model.conv_pre_mag.weight.device
+        return self
+
+    def __call__(self, audio, orig_sampling_rate):
+        with torch.no_grad():
+            # audio, orig_sampling_rate = torchaudio.load(inp_path)
+            # audio = audio.to(self.device)
+            audio = aF.resample(audio, orig_freq=orig_sampling_rate, new_freq=self.h.hr_sampling_rate)
+            amp_nb, pha_nb, com_nb = amp_pha_stft(audio, self.h.n_fft, self.h.hop_size, self.h.win_size)
+            amp_wb_g, pha_wb_g, com_wb_g = self.model(amp_nb, pha_nb)
+            audio_hr_g = amp_pha_istft(amp_wb_g, pha_wb_g, self.h.n_fft, self.h.hop_size, self.h.win_size)
+            # sf.write(opt_path, audio_hr_g.squeeze().cpu().numpy(), self.h.hr_sampling_rate, 'PCM_16')
+            return audio_hr_g.squeeze().cpu().numpy(), self.h.hr_sampling_rate
diff --git a/tools/cmd-denoise.py b/tools/cmd-denoise.py
new file mode 100644
index 0000000000000000000000000000000000000000..bbf68476405be4e9fa2152095e3ed737a0142844
--- /dev/null
+++ b/tools/cmd-denoise.py
@@ -0,0 +1,38 @@
+import os
+import argparse
+import traceback
+
+from modelscope.pipelines import pipeline
+from modelscope.utils.constant import Tasks
+from tqdm import tqdm
+
+path_denoise = "tools/denoise-model/speech_frcrn_ans_cirm_16k"
+path_denoise = path_denoise if os.path.exists(path_denoise) else "damo/speech_frcrn_ans_cirm_16k"
+ans = pipeline(Tasks.acoustic_noise_suppression, model=path_denoise)
+
+
+def execute_denoise(input_folder, output_folder):
+    os.makedirs(output_folder, exist_ok=True)
+    # print(input_folder)
+    # print(list(os.listdir(input_folder).sort()))
+    for name in tqdm(os.listdir(input_folder)):
+        try:
+            ans("%s/%s" % (input_folder, name), output_path="%s/%s" % (output_folder, name))
+        except:
+            traceback.print_exc()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "-i", "--input_folder", type=str, required=True, help="Path to the folder containing WAV files."
+    )
+    parser.add_argument("-o", "--output_folder", type=str, required=True, help="Output folder to store transcriptions.")
+    parser.add_argument(
+        "-p", "--precision", type=str, default="float16", choices=["float16", "float32"], help="fp16 or fp32"
+    )  # 还没接入
+    cmd = parser.parse_args()
+    execute_denoise(
+        input_folder=cmd.input_folder,
+        output_folder=cmd.output_folder,
+    )
diff --git a/tools/denoise-model/.gitignore b/tools/denoise-model/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/tools/denoise-model/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/tools/i18n/i18n.py b/tools/i18n/i18n.py
new file mode 100644
index 0000000000000000000000000000000000000000..4cd123f3e53b69e56ef251aa573c3c2ad186205b
--- /dev/null
+++ b/tools/i18n/i18n.py
@@ -0,0 +1,41 @@
+import json
+import locale
+import os
+
+I18N_JSON_DIR: os.PathLike = os.path.join(os.path.dirname(os.path.relpath(__file__)), "locale")
+
+
+def load_language_list(language):
+    with open(os.path.join(I18N_JSON_DIR, f"{language}.json"), "r", encoding="utf-8") as f:
+        language_list = json.load(f)
+    return language_list
+
+
+def scan_language_list():
+    language_list = []
+    for name in os.listdir(I18N_JSON_DIR):
+        if name.endswith(".json"):
+            language_list.append(name.split(".")[0])
+    return language_list
+
+
+class I18nAuto:
+    def __init__(self, language=None):
+        if language in ["Auto", None]:
+            language = locale.getdefaultlocale()[0]
+            # getlocale can't identify the system's language ((None, None))
+        if not os.path.exists(os.path.join(I18N_JSON_DIR, f"{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 "Use Language: " + self.language
+
+
+if __name__ == "__main__":
+    i18n = I18nAuto(language="en_US")
+    print(i18n)
diff --git a/tools/i18n/locale/en_US.json b/tools/i18n/locale/en_US.json
new file mode 100644
index 0000000000000000000000000000000000000000..42c8704457cd1ca4c4d8a54ded2b6201ca891907
--- /dev/null
+++ b/tools/i18n/locale/en_US.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb): Best choice for dual-channel reverberation, cannot remove single-channel reverberation;",
+    "(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。": "(234)DeEcho: Removes delay effects. Aggressive mode removes more thoroughly than Normal mode. DeReverb additionally removes reverberation, can remove mono reverberation, but does not clean heavily high-frequency plate reverberation.",
+    "*实验/模型名": "*Experiment/model name",
+    "*文本标注文件": "*Text labelling file",
+    "*训练集音频文件目录": "*Audio dataset folder",
+    "*请上传并填写参考信息": "*Please upload and fill reference information",
+    "*请填写需要合成的目标文本和语种模式": "*Please fill in the target text and language mode for synthesis",
+    ".限制范围越小判别效果越好。": "Less Multilingual is better",
+    "1-GPT-SoVITS-TTS": "1-GPT-SOVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. The DeEcho-DeReverb model's processing time is nearly twice that of the other two DeEcho models.",
+    "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；": "1. Preserve Vocals: Choose this option for audio without harmonies, as it better retains the main vocal compared to the HP5 model. This option includes two built-in models, HP2 and HP3. HP3 may slightly let through some accompaniment but retains the main vocal slightly better than HP2.",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-Voice Changer",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverb Model is slow;",
+    "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；": "2. Keep Only Main Vocal: Choose this option for audio with harmonies, as it may slightly reduce the main vocal. Includes one built-in HP5 model;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. Personal Recommendation for the cleanest configuration: First use MDX-Net followed by DeEcho-Aggressive",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Reverberation and delay removal model(by FoxJoy):",
+    "ASR 模型": "ASR model",
+    "ASR 模型尺寸": "ASR model size",
+    "ASR 语言设置": "ASR language",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT Training: Model Weights saved in GPT_weights/",
+    "GPT模型列表": "GPT weight list",
+    "GPT训练": "GPT Training",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT sampling parameters (not too low when there's no reference text. Use default if unsure):",
+    "GPU卡号,只能填1个整数": "GPU number, can only input ONE integer",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU number is separated by -, each GPU will run one process ",
+    "LoRA秩": "LoRA Rank",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "Missing Pretrained SoVITS V3 Model, Cannot Load LoRA Weights",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS Training: Model Weights saved in SoVITS_weights/",
+    "SoVITS模型列表": "SoVITS weight list",
+    "SoVITS训练": "SoVITS Training",
+    "TTS推理WebUI": "TTS Inference WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5 WebUI (Vocal Separation/Deecho/Dereverb)",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix: proportion of normalized audio merged into dataset",
+    "batch_size": "Batch Size",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: FO hop size, the smaller the value, the higher the accuracy）",
+    "max:归一化后最大值多少": "Loudness multiplier after normalized",
+    "max_sil_kept:切完后静音最多留多长": "Maximum length for silence to be kept",
+    "min_interval:最短切割间隔": "Minumum interval for audio cutting",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: the minimum length of each segment. If the first segment is too short, it will be concatenated with the next segment until it exceeds this value",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "Noise gate threshold (loudness below this value will be treated as noise",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3 does not support this mode currently, using it will cause an error.",
+    "v3输出如果觉得闷可以试试开超分": "For V3 model, if generated audio sounds somewhat muffled, try enable audio super-resolution.",
+    "不切": "No slice",
+    "中文": "Chinese",
+    "中文教程文档": "Chinese Tutorial",
+    "中英混合": "Chinese-English Mixed",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Primary Reference Audio (Please upload reference audio within 3-10 seconds, exceeding this limit will cause an error!)",
+    "主参考音频的文本": "Text of Primary Reference Audio",
+    "主参考音频的语种": "Language of Primary Reference Audio",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "Multiple audio files can also be imported. If a folder path exists, this input is ignored.",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Batch processing for vocal and instrumental separation, using the UVR5 model.",
+    "人声分离WebUI": "Vocal Separation WebUI",
+    "人声提取激进程度": "Vocal extraction aggressiveness",
+    "以下文件或文件夹不存在": "No Such File or Folder",
+    "以下模型不存在:": "No Such Model:",
+    "伴奏人声分离&去混响&去回声": "Vocals/Accompaniment Separation & Reverberation Removal",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Super-Resolution Model Not Found. Please follow the tutorial to download the model file if you want to use it.",
+    "使用无参考文本模式时建议使用微调的GPT": "Recommended to use a Finetune-GPT when using Prompt-Free Mode.",
+    "保存频率save_every_epoch": "Save frequency (save_every_epoch):",
+    "保持随机": "Keep Random",
+    "关闭": "Close ",
+    "凑50字一切": "Slice per 50 characters",
+    "凑四句一切": "Slice once every 4 sentences",
+    "分桶处理模式已关闭": "Bucket Processing Mode Disabled",
+    "分桶处理模式已开启": "Bucket Processing Mode Enabled",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "Segmented Return Mode does not support Bucket Processing, Bucket Processing Disabled automatically",
+    "分段返回模式已开启": "Segmented Return Mode Enabled",
+    "分段间隔(秒)": "Segment Interval (Seconds)",
+    "分段间隔过小，已自动设置为0.01": "Segment Interval too short, automatically set to 0.01",
+    "切分": "Segmentation",
+    "切分后文本": "Inference Text After Segmentation",
+    "切分后的子音频的输出根目录": "Audio slicer output folder",
+    "切分文本": "Segment Text",
+    "切割使用的进程数": "CPU threads used for audio slicing",
+    "刷新模型路径": "refreshing model paths",
+    "前端处理后的文本(每句):": "Processed text from the frontend (per sentence):",
+    "前置数据集获取工具": "Fetch Datasets",
+    "占用中": " Occupying",
+    "去混响/去延迟，附：": "Dereverberation/Delay Removal, including:",
+    "参考音频在3~10秒范围外，请更换！": "Reference audio is outside the 3-10 second range, please choose another one!",
+    "参考音频的文本": "Text for reference audio",
+    "参考音频的语种": "Language for reference audio",
+    "句间停顿秒数": "Pause Duration between Sentences (Seconds)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Optional: Upload multiple reference audio files by dragging and dropping them (recommended to be of the same gender), and average their tone. If this option is left blank, the tone will be controlled by the single reference audio on the left. If fine-tuning the model, it is recommended that all reference audio files have tones within the fine-tuning training set; the pretrained model can be ignored.",
+    "合成语音": "Start inference",
+    "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)。": "An example of a valid folder path format: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (simply copy the address from the file manager's address bar).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "Support for Phoneme Conversion, Manual Phoneme Editing, and Step-by-Step Speech Synthesis will be added in the future.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "If reference audio is not clear or unsure what to write, enable this option to ignore the reference text.",
+    "启用并行推理版本": "Enable Parallel Inference Version",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Please fill in the segmented audio files' directory! The full path of the audio file = the directory concatenated with the filename corresponding to the waveform in the list file (not the full path). If left blank, the absolute full path in the .list file will be used.",
+    "多语种混合": "Multilingual Mixed",
+    "多语种混合(粤语)": "Multilingual Mixed(Yue)",
+    "失败": " Failed",
+    "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.": "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 for details.",
+    "实际输入的参考文本:": "Actual Input Reference Text:",
+    "实际输入的目标文本(切句后):": "Actual Input Target Text (after sentence segmentation):",
+    "实际输入的目标文本(每句):": "Actual Input Target Text (per sentence):",
+    "实际输入的目标文本:": "Actual Input Target Text:",
+    "导出文件格式": "Export file format",
+    "已关闭": " is Closed",
+    "已完成": " Finished",
+    "已开启": " is Opened",
+    "并行推理": "Parallel Inference",
+    "并行推理模式已关闭": "Parallel Inference Mode Disabled",
+    "并行推理模式已开启": "Parallel Inference Mode Enabled",
+    "开启": "Open ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Enable no reference mode. If you don't fill 'Text for reference audio', no reference mode will be enabled.",
+    "微调训练": "Fine-Tuning",
+    "怎么切": "How to slice the sentence",
+    "总训练轮数total_epoch": "Total training epochs (total_epoch):",
+    "总训练轮数total_epoch，不建议太高": "Total epochs, do not increase to a value that is too high",
+    "指定输出主人声文件夹": "Specify the output folder for vocals:",
+    "指定输出非主人声文件夹": "Specify the output folder for accompaniment:",
+    "按中文句号。切": "Slice by Chinese punct",
+    "按标点符号切": "Slice by every punct",
+    "按英文句号.切": "Slice by English punct",
+    "推理": "Inference",
+    "推理设置": "Inference Settings",
+    "提取文本Bert特征": "Extract Text BERT Features",
+    "数据分桶(并行推理时会降低一点计算量)": "Data Bucketing (Reduces Computation Cost in Parallel Inference)",
+    "数据类型精度": "Computing precision",
+    "文本分词与特征提取": "Tokenization & BERT Feature Extraction",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Text Segmentation Tool. Very long text may not yield good synthesis results, so Segmentation is Recommended. Synthesis will be performed based on line breaks and then concatenated.",
+    "文本模块学习率权重": "Text model learning rate weighting",
+    "施工中，请静候佳音": "In construction, please wait",
+    "日文": "Japanese",
+    "日英混合": "Japanese-English Mixed",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Save only the latest weight file to save disk space",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Save a small final model to the 'weights' folder at each save point:",
+    "是否开启DPO训练选项(实验性)": "Enable DPO Training (Experimental)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Adjust the speech rate and tone of the last synthesis result to prevent randomness.",
+    "显卡信息": "GPU Information",
+    "未下载模型": "Model Not Downloaded",
+    "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.": "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.",
+    "标注文件路径 (含文件后缀 *.list)": "Label File Path (with file extension *.list)",
+    "模型": "Model",
+    "模型分为三类：": "Models are categorized into three types:",
+    "模型切换": "Model switch",
+    "每张显卡的batch_size": "Batch size per GPU:",
+    "版本": "Version",
+    "粤英混合": "Yue-English Mixed",
+    "粤语": "Yue",
+    "终止合成": "Terminate Synthesis",
+    "缺少Hubert数据集": "Missing Hubert Dataset",
+    "缺少语义数据集": "Missing Semantics Dataset",
+    "缺少音素数据集": "Missing Phoneme Dataset",
+    "缺少音频数据集": "Missing Audio Dataset",
+    "英文": "English",
+    "训练集格式化一键三连": "Training Set One-Click Formatting",
+    "训练集格式化工具": "Dataset Formatting Tool",
+    "语义Token提取": "Semantics Token Extraction",
+    "语速": "Speech rate",
+    "语速调整，高为更快": "Adjust speech rate, higher for faster",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Speech Rate Adjustment does not support Bucket Processing, Bucket Processing Disabled automatically",
+    "语音切分": "Speech Slicing",
+    "语音切分工具": "Speech Slicing Tool",
+    "语音文本校对标注工具": "Speech-to-Text Proofreading Tool",
+    "语音自监督特征提取": "Speech SSL Feature Extraction",
+    "语音识别": "Speech Recognition",
+    "语音识别工具": "Speech Recognition Tool",
+    "语音降噪": "Speech Denoising",
+    "语音降噪工具": "Speech Denoising Tool",
+    "请上传3~10秒内参考音频，超过会报错！": "Please upload a reference audio within the 3-10 second range; if it exceeds this duration, it will raise errors.",
+    "请上传参考音频": "Please Upload the Reference Audio",
+    "请填入推理文本": "Please Fill in the Terget Text",
+    "请填入正确的List路径": "Please Fill in the Correct List Path",
+    "请填入正确的音频文件夹路径": "Please Fill in the Correct Audio Folder Path",
+    "请输入有效文本": "Please enter valid text.",
+    "路径不存在,使用默认配置": "Paths Not Found, Using Default Config",
+    "路径不能为空": "Expected No Empty Path",
+    "路径错误": "Path Error",
+    "转换": "Convert",
+    "辅参考音频(可选多个，或不选)": "Secondary Reference Audio (Multiple Optional, or None)",
+    "输入待处理音频文件夹路径": "Enter the path of the audio folder to be processed:",
+    "输入文件夹路径": "Input folder path",
+    "输入路径不存在": "Input Path Not Found",
+    "输入路径存在但不可用": "Input Path Exists but Unavailable",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "output folder (logs/{experiment name}) should have files and folders starts with 23456.",
+    "输出信息": "Output information",
+    "输出文件夹路径": "Output folder path",
+    "输出的语音": "Inference Result",
+    "运行中": " Running",
+    "进度": "Progress",
+    "进程已终止": " Process Terminated",
+    "进程输出信息": " Process Output Information",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Choose the models from SoVITS_weights and GPT_weights. The default one is a pretrain, so you can experience zero shot TTS.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Sampling Steps: If feel noisy, try increasing, if feel slow, try decreasing",
+    "重复惩罚": "Repetition Penalty",
+    "随机种子": "Random Seed",
+    "需先终止才能开启下一次任务": "Please Terminate First to Start Next Task",
+    "需要合成的切分前文本": "Inference Text Before Segmentation",
+    "需要合成的文本": "Inference text",
+    "需要合成的文本的语种": "Language of the Inference Text",
+    "需要合成的语种": "Inference text language",
+    "韩文": "Korean",
+    "韩英混合": "Korean-English Mixed",
+    "音频加载失败": "Failed to Load Audio",
+    "音频文件不存在，跳过：": "Audio File Not Found, Skipping: ",
+    "音频标注WebUI": "Audio Labeling WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "Audio slicer input (file or folder)",
+    "音频超分中": "Running Audio Super-Resolution",
+    "预训练GPT模型路径": "Pretrained GPT Model Path",
+    "预训练SSL模型路径": "Pretrained SSL Model Path",
+    "预训练SoVITS-D模型路径": "Pretrained SoVITS-D Model Path",
+    "预训练SoVITS-G模型路径": "Pretrained SoVITS-G Model Path",
+    "预训练中文BERT模型路径": "Pretrained Chinese BERT Model Path"
+}
diff --git a/tools/i18n/locale/es_ES.json b/tools/i18n/locale/es_ES.json
new file mode 100644
index 0000000000000000000000000000000000000000..3e99ba8b08cf6b09efc5ca24c7e5b890a6674d95
--- /dev/null
+++ b/tools/i18n/locale/es_ES.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net (onnx_dereverb): reverberación estéreo, la mejor opción; no puede eliminar reverberación mono",
+    "(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。": "(234)DeEcho: Eliminar el efecto de retardo. Aggressive elimina más que Normal, DeReverb elimina reverberación adicional, puede eliminar reverberación mono, pero no limpia bien la reverberación de placa de alta frecuencia",
+    "*实验/模型名": "*Nombre del experimento/modelo",
+    "*文本标注文件": "*Archivo de etiquetado de texto",
+    "*训练集音频文件目录": "*Directorio de archivos de audio de entrenamiento",
+    "*请上传并填写参考信息": "*Por favor, suba y complete la información de referencia",
+    "*请填写需要合成的目标文本和语种模式": "*Por favor, complete el texto objetivo a sintetizar y el modo de idioma",
+    ".限制范围越小判别效果越好。": ".Cuanto más pequeño sea el rango, mejor será el efecto de discriminación.",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. El modelo DeEcho-DeReverb tarda casi el doble que los otros dos modelos DeEcho",
+    "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；": "1. Retener voz principal: seleccione este para audio sin coros, retiene mejor la voz principal que HP5. Incluye dos modelos, HP2 y HP3; HP3 puede filtrar ligeramente el acompañamiento pero retiene mejor la voz principal que HP2",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-Cambio de voz",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. El modelo MDX-Net-Dereverb es bastante lento",
+    "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；": "2. Solo retener voz principal: seleccione este para audio con coros, puede debilitar la voz principal. Incluye un modelo HP5",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. La configuración más limpia recomendada es primero MDX-Net, luego DeEcho-Aggressive",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Modelos de eliminación de reverberación y retardo (por FoxJoy)",
+    "ASR 模型": "Modelo ASR",
+    "ASR 模型尺寸": "Tamaño del modelo ASR",
+    "ASR 语言设置": "Configuración del idioma ASR",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "Entrenamiento de GPT: los archivos de pesos del modelo están en GPT_weights/",
+    "GPT模型列表": "Lista de modelos GPT",
+    "GPT训练": "Entrenamiento de GPT",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "Parámetros de muestreo de GPT (no demasiado bajos cuando no hay texto de referencia. Use los valores por defecto si no está seguro):",
+    "GPU卡号,只能填1个整数": "Número de tarjeta GPU, solo se puede ingresar un número entero",
+    "GPU卡号以-分割，每个卡号一个进程": "Número de tarjeta GPU separado por '-', cada número de tarjeta es un proceso",
+    "LoRA秩": "Rango de LoRA",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "Falta el modelo base de SoVITS V3, no se pueden cargar los pesos de LoRA correspondientes",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "Entrenamiento de SoVITS: los archivos de pesos del modelo están en SoVITS_weights/",
+    "SoVITS模型列表": "Lista de modelos SoVITS",
+    "SoVITS训练": "Entrenamiento de SoVITS",
+    "TTS推理WebUI": "WebUI de inferencia TTS",
+    "UVR5人声伴奏分离&去混响去延迟工具": "Herramienta de separación de voz y acompañamiento UVR5 y eliminación de reverberación y retardo",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix: proporción de mezcla de audio normalizado que entra",
+    "batch_size": "Tamaño de lote",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: cómo calcular la curva de volumen, cuanto más pequeño, mayor precisión pero mayor carga computacional (mayor precisión no significa mejor rendimiento)",
+    "max:归一化后最大值多少": "max: valor máximo después de la normalización",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept: duración máxima del silencio después del corte",
+    "min_interval:最短切割间隔": "min_interval: intervalo mínimo de corte",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: longitud mínima de cada segmento; si el primer segmento es demasiado corto, se une al siguiente hasta superar este valor",
+    "temperature": "temperatura",
+    "threshold:音量小于这个值视作静音的备选切割点": "umbral: puntos de corte alternativos considerados como silencio si el volumen es menor que este valor",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3 no es compatible con este modo actualmente y su uso generará un error.",
+    "v3输出如果觉得闷可以试试开超分": "Si la salida de V3 parece aburrida, puedes intentar activar la superresolución",
+    "不切": "No cortar",
+    "中文": "Chino",
+    "中文教程文档": "Documentación del tutorial en chino",
+    "中英混合": "Chino e inglés mezclados",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Audio de referencia principal (Por favor, suba un audio de referencia de entre 3 y 10 segundos, si supera este límite se producirá un error)",
+    "主参考音频的文本": "Texto del audio de referencia principal",
+    "主参考音频的语种": "Idioma del audio de referencia principal",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "También se pueden ingresar archivos de audio por lotes, seleccionar uno, prioridad para leer carpetas",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Procesamiento por lotes de separación de voz y acompañamiento utilizando el modelo UVR5",
+    "人声分离WebUI": "WebUI de separación de voces",
+    "人声提取激进程度": "Nivel de agresividad en la extracción de voz",
+    "以下文件或文件夹不存在": "No Existe Tal Archivo o Carpeta",
+    "以下模型不存在:": "No Existe tal Modelo:",
+    "伴奏人声分离&去混响&去回声": "Separación de acompañamiento y voz principal y eliminación de reverberación y eco",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "No has descargado los parámetros del modelo de superresolución, por lo que no se realizará la superresolución. Si deseas habilitarla, sigue el tutorial y descarga los archivos necesarios",
+    "使用无参考文本模式时建议使用微调的GPT": "Se recomienda usar un GPT ajustado cuando se use el modo sin texto de referencia.",
+    "保存频率save_every_epoch": "Frecuencia de guardado (cada epoch)",
+    "保持随机": "Mantener aleatorio",
+    "关闭": "Cerrar ",
+    "凑50字一切": "Todo para alcanzar las 50 palabras",
+    "凑四句一切": "Completa cuatro oraciones para rellenar todo",
+    "分桶处理模式已关闭": "Modo de procesamiento por lotes deshabilitado",
+    "分桶处理模式已开启": "Modo de procesamiento por lotes habilitado",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "El modo de retorno segmentado no es compatible con el procesamiento por lotes, se ha deshabilitado automáticamente",
+    "分段返回模式已开启": "Modo de retorno segmentado habilitado",
+    "分段间隔(秒)": "Intervalo de segmentación (segundos)",
+    "分段间隔过小，已自动设置为0.01": "El intervalo de segmentación es demasiado pequeño, se ha ajustado automáticamente a 0.01",
+    "切分": "Segmentación",
+    "切分后文本": "Texto después de la segmentación",
+    "切分后的子音频的输出根目录": "Directorio raíz de salida de los sub-audios después de la división",
+    "切分文本": "Segmentar texto",
+    "切割使用的进程数": "Número de procesos utilizados para la división",
+    "刷新模型路径": "Actualizar la ruta del modelo",
+    "前端处理后的文本(每句):": "Texto después del procesamiento previo (por frase):",
+    "前置数据集获取工具": "Herramienta de adquisición de conjunto de datos previo",
+    "占用中": " En uso",
+    "去混响/去延迟，附：": "Eliminación de reverberación/retardo, incluye:",
+    "参考音频在3~10秒范围外，请更换！": "El audio de referencia está fuera del rango de 3 a 10 segundos, ¡por favor cámbielo!",
+    "参考音频的文本": "Texto de referencia del audio",
+    "参考音频的语种": "Idioma del audio de referencia",
+    "句间停顿秒数": "Segundos de pausa entre frases",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Opcional: Sube varios archivos de audio de referencia arrastrándolos y soltándolos (se recomienda que sean del mismo género) y promedia sus tonos. Si esta opción se deja en blanco, el tono será controlado por el único audio de referencia a la izquierda. Si se está afinando el modelo, se recomienda que todos los archivos de audio de referencia tengan tonos dentro del conjunto de entrenamiento de ajuste fino; se puede ignorar el modelo preentrenado.",
+    "合成语音": "Síntesis de voz",
+    "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)。": "Ejemplo de formato de ruta de carpeta válida: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (simplemente copie desde la barra de direcciones del administrador de archivos).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "Se añadirá soporte para conversión de fonemas, edición manual de fonemas y síntesis de voz por pasos en el futuro.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Si el audio de referencia no es claro o no sabe qué escribir, habilite esta opción para ignorar el texto de referencia.",
+    "启用并行推理版本": "Habilitar versión de inferencia paralela",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Ingrese el directorio donde se encuentran los audios después de la división. La ruta completa de los archivos de audio leídos = este directorio + nombre de archivo correspondiente en el archivo .list (no la ruta completa). Si se deja en blanco, se utilizará la ruta completa del archivo .list.",
+    "多语种混合": "Mezcla de varios idiomas",
+    "多语种混合(粤语)": "Mezcla Multilingüe (Cantonés)",
+    "失败": " Fallido",
+    "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.": "Si no acepta estos términos, no puede utilizar ni hacer referencia a ningún código o archivo dentro del paquete de software. Consulte el archivo LICENSE en el directorio raíz para obtener más detalles.",
+    "实际输入的参考文本:": "Texto de referencia realmente ingresado:",
+    "实际输入的目标文本(切句后):": "Texto objetivo realmente ingresado (después de dividir en frases):",
+    "实际输入的目标文本(每句):": "Texto objetivo realmente ingresado (por frase):",
+    "实际输入的目标文本:": "Texto objetivo realmente ingresado:",
+    "导出文件格式": "Formato de archivo de exportación",
+    "已关闭": " Desactivado",
+    "已完成": " Completado",
+    "已开启": " Activado",
+    "并行推理": "Inferencia paralela",
+    "并行推理模式已关闭": "Modo de inferencia paralela deshabilitado",
+    "并行推理模式已开启": "Modo de inferencia paralela habilitado",
+    "开启": "Activar ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Habilitar el modo sin texto de referencia. No llenar el texto de referencia también lo habilita.",
+    "微调训练": "Entrenamiento de ajuste fino",
+    "怎么切": "Cómo cortar",
+    "总训练轮数total_epoch": "Número total de épocas de entrenamiento",
+    "总训练轮数total_epoch，不建议太高": "Número total de épocas de entrenamiento, no se recomienda demasiado alto",
+    "指定输出主人声文件夹": "Especificar carpeta de salida de voz principal",
+    "指定输出非主人声文件夹": "Especificar carpeta de salida de no voz principal",
+    "按中文句号。切": "Cortar según puntos en chino",
+    "按标点符号切": "Cortar según los signos de puntuación",
+    "按英文句号.切": "Cortar por puntos en inglés.",
+    "推理": "Inferencia",
+    "推理设置": "Configuración de inferencia",
+    "提取文本Bert特征": "Extraer características de texto con BERT",
+    "数据分桶(并行推理时会降低一点计算量)": "Agrupación de datos (Reduce el costo computacional en inferencia paralela)",
+    "数据类型精度": "precisión del tipo de datos",
+    "文本分词与特征提取": "Segmentación de texto y extracción de características",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Herramienta de segmentación de texto. Un texto demasiado largo puede no producir buenos resultados, por lo que se recomienda segmentarlo. La síntesis se realizará por separado según los saltos de línea y luego se unirá.",
+    "文本模块学习率权重": "Peso de la tasa de aprendizaje del módulo de texto",
+    "施工中，请静候佳音": "En construcción, por favor espere pacientemente",
+    "日文": "Japonés",
+    "日英混合": "Mezcla de japonés e inglés",
+    "是否仅保存最新的权重文件以节省硬盘空间": "¿Guardar solo el último archivo de pesos más reciente para ahorrar espacio en disco?",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "¿Guardar el modelo final pequeño en la carpeta de pesos en cada punto de guardado?",
+    "是否开启DPO训练选项(实验性)": "¿Habilitar la opción de entrenamiento dpo (experimental)?",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "¿Ajustar directamente la velocidad del habla y el tono del último resultado de síntesis? Para prevenir la aleatoriedad.",
+    "显卡信息": "Información de la tarjeta gráfica",
+    "未下载模型": "Modelo no descargado",
+    "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.": "Este software es de código abierto bajo la licencia MIT. El autor no tiene control sobre el software. El usuario que lo utilice o distribuya, y el que genere sonidos a partir del software, asume toda la responsabilidad.",
+    "标注文件路径 (含文件后缀 *.list)": "Ruta del archivo de anotaciones (con extensión *.list)",
+    "模型": "Modelo",
+    "模型分为三类：": "Los modelos se dividen en tres categorías:",
+    "模型切换": "Cambio de modelo",
+    "每张显卡的batch_size": "Tamaño de lote por tarjeta gráfica",
+    "版本": "Versión",
+    "粤英混合": "Mezcla Cantonés-Inglés",
+    "粤语": "Cantonés",
+    "终止合成": "Terminar síntesis",
+    "缺少Hubert数据集": "Falta el Conjunto de Datos de Hubert",
+    "缺少语义数据集": "Falta el Conjunto de Datos Semánticos",
+    "缺少音素数据集": "Falta el Conjunto de Datos de Fonemas",
+    "缺少音频数据集": "Falta el Conjunto de Datos de Audio",
+    "英文": "Inglés",
+    "训练集格式化一键三连": "Formato del conjunto de entrenamiento en un solo paso",
+    "训练集格式化工具": "Herramienta de formateo del conjunto de datos de entrenamiento",
+    "语义Token提取": "Extracción de tokens semánticos",
+    "语速": "Velocidad de habla",
+    "语速调整，高为更快": "Ajustar la velocidad de habla, más alta para más rápido",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "El ajuste de velocidad de voz no es compatible con el procesamiento por lotes, se ha deshabilitado automáticamente",
+    "语音切分": "Segmentación de voz",
+    "语音切分工具": "Herramienta de división de voz",
+    "语音文本校对标注工具": "Herramienta de corrección y anotación de texto de voz",
+    "语音自监督特征提取": "Extracción de características de voz con auto-supervisión",
+    "语音识别": "Reconocimiento de voz",
+    "语音识别工具": "Herramienta de reconocimiento de voz",
+    "语音降噪": "Reducción de ruido en la voz",
+    "语音降噪工具": "Herramienta de reducción de ruido de voz",
+    "请上传3~10秒内参考音频，超过会报错！": "Por favor, suba un audio de referencia de entre 3 y 10 segundos, ¡más de eso causará un error!",
+    "请上传参考音频": "Por Favor, Suba el Audio de Referencia",
+    "请填入推理文本": "Por Favor, Ingrese el Texto Objetivo",
+    "请填入正确的List路径": "Por Favor, Introduzca la Ruta Correcta de la Lista",
+    "请填入正确的音频文件夹路径": "Por Favor, Introduzca la Ruta Correcta de la Carpeta de Audio",
+    "请输入有效文本": "Por favor, introduzca un texto válido",
+    "路径不存在,使用默认配置": "Ruta no encontrada, usando configuración predeterminada",
+    "路径不能为空": "Se Espera que la Ruta No Esté Vacía",
+    "路径错误": "Error de Ruta",
+    "转换": "Convertir",
+    "辅参考音频(可选多个，或不选)": "Audio de referencia secundario (Opcional, se pueden seleccionar varios o ninguno)",
+    "输入待处理音频文件夹路径": "Ingrese la ruta de la carpeta de audio a procesar",
+    "输入文件夹路径": "Ingrese la ruta de la carpeta",
+    "输入路径不存在": "La ruta de entrada no existe",
+    "输入路径存在但不可用": "La ruta de entrada existe pero no es accesible",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "Debe haber archivos y carpetas que comiencen con 23456 en el directorio logs/nombre del experimento",
+    "输出信息": "Información de salida",
+    "输出文件夹路径": "Ruta de la carpeta de salida",
+    "输出的语音": "Audio de salida",
+    "运行中": " En ejecución",
+    "进度": "Progreso",
+    "进程已终止": " Proceso terminado",
+    "进程输出信息": " Información de salida del proceso",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Seleccione el modelo almacenado en SoVITS_weights y GPT_weights después del entrenamiento. Uno de ellos es el modelo base, útil para experimentar con TTS de 5 segundos sin entrenamiento.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Pasos de muestreo: si se siente ruidoso, intente aumentarlo; si es lento, intente reducirlo",
+    "重复惩罚": "Penalización por repetición",
+    "随机种子": "Semilla aleatoria",
+    "需先终止才能开启下一次任务": " Debe terminarse antes de iniciar la siguiente tarea",
+    "需要合成的切分前文本": "Texto antes de la segmentación para la síntesis",
+    "需要合成的文本": "Texto a sintetizar",
+    "需要合成的文本的语种": "Idioma del texto a sintetizar",
+    "需要合成的语种": "Idioma para la síntesis",
+    "韩文": "Coreano",
+    "韩英混合": "Mezcla Coreano-Inglés",
+    "音频加载失败": "Error al Cargar el Audio",
+    "音频文件不存在，跳过：": "Archivo de audio no encontrado, omitiendo: ",
+    "音频标注WebUI": "WebUI de etiquetado de audio",
+    "音频自动切分输入路径，可文件可文件夹": "Ruta de entrada para la división automática de audio, puede ser un archivo o una carpeta",
+    "音频超分中": "Superresolución de audio en proceso",
+    "预训练GPT模型路径": "Ruta del modelo GPT preentrenado",
+    "预训练SSL模型路径": "Ruta del modelo SSL preentrenado",
+    "预训练SoVITS-D模型路径": "Ruta del modelo SoVITS-D preentrenado",
+    "预训练SoVITS-G模型路径": "Ruta del modelo SoVITS-G preentrenado",
+    "预训练中文BERT模型路径": "Ruta del modelo BERT en Chino preentrenado"
+}
diff --git a/tools/i18n/locale/fr_FR.json b/tools/i18n/locale/fr_FR.json
new file mode 100644
index 0000000000000000000000000000000000000000..dd9425046adfc5ec44ea04e516598c27ca42ec69
--- /dev/null
+++ b/tools/i18n/locale/fr_FR.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1) MDX-Net (onnx_dereverb) : C'est le meilleur choix pour la réverbération à deux canaux, mais il ne peut pas éliminer la réverbération à un seul canal;",
+    "(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。": "(234)DeEcho : Supprime les effets de délai. Aggressive est plus exhaustif que Normal dans la suppression, DeReverb élimine également la réverbération, peut supprimer la réverbération monocanal, mais n'élimine pas complètement la réverbération de plaque à haute fréquence.",
+    "*实验/模型名": "*Nom de l'expérience/modèle",
+    "*文本标注文件": "*Fichier d'annotation de texte",
+    "*训练集音频文件目录": "*Répertoire des fichiers audio d'entraînement",
+    "*请上传并填写参考信息": "*Veuillez télécharger et remplir les informations de référence",
+    "*请填写需要合成的目标文本和语种模式": "*Veuillez saisir le texte cible à synthétiser et le mode de langue.",
+    ".限制范围越小判别效果越好。": "Moins il y a de langues, mieux c'est",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. Le temps de traitement du modèle DeEcho-DeReverb est presque le double de celui des deux autres modèles DeEcho;",
+    "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；": "1. Préserver les voix : Choisissez cette option pour les audio sans harmonie, car elle conserve mieux la voix principale par rapport au modèle HP5. Deux modèles intégrés, HP2 et HP3, sont disponibles. HP3 peut légèrement laisser passer l'accompagnement mais conserve la voix principale un peu mieux que HP2;",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-Modification de la voix",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. Le modèle MDX-Net-Dereverb est assez lent;",
+    "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；": "2. Conserver uniquement la voix principale : Choisissez cette option pour les audio avec harmonie, car elle peut affaiblir la voix principale. Un modèle HP5 intégré est disponible;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. La configuration la plus propre que je recommande est d'utiliser d'abord MDX-Net, puis DeEcho-Aggressive.",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Modèle de suppression de réverbération et de retard (par FoxJoy) :",
+    "ASR 模型": "Modèle ASR",
+    "ASR 模型尺寸": "Taille du modèle ASR",
+    "ASR 语言设置": "Paramètres de langue ASR",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "Entraînement GPT : les poids du modèle sont dans GPT_weights/",
+    "GPT模型列表": "Liste des modèles GPT",
+    "GPT训练": "Entraînement GPT",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "Paramètres d'échantillonnage de GPT (ne pas mettre trop bas lorsqu'il n'y a pas de texte de référence. Utilisez les valeurs par défaut si vous n'êtes pas sûr):",
+    "GPU卡号,只能填1个整数": "Numéro de carte GPU, ne peut contenir qu'un seul entier",
+    "GPU卡号以-分割，每个卡号一个进程": "Numéro de carte GPU séparé par des tirets, un processus par numéro de carte",
+    "LoRA秩": "Rang LoRA",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "Modèle de base SoVITS V3 manquant, impossible de charger les poids LoRA correspondants",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "Entraînement SoVITS : les poids du modèle sont dans SoVITS_weights/",
+    "SoVITS模型列表": "Liste des modèles SoVITS",
+    "SoVITS训练": "Entraînement SoVITS",
+    "TTS推理WebUI": "Interface Web d'inférence TTS",
+    "UVR5人声伴奏分离&去混响去延迟工具": "Outil UVR5 de séparation voix/accompagnement & suppression de réverbération et de latence",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix: proportion d'audio normalisé mélangé",
+    "batch_size": "Taille de lot",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: comment calculer la courbe de volume, plus petit pour une précision plus élevée mais une charge de calcul plus élevée (ce n'est pas une meilleure précision)",
+    "max:归一化后最大值多少": "max: valeur maximale après normalisation",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept: durée maximale de silence après la coupe",
+    "min_interval:最短切割间隔": "min_interval: intervalle de coupe minimum",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:longueur minimale de chaque segment ; si le premier segment est trop court, il est concaténé avec les segments suivants jusqu'à ce que la longueur dépasse cette valeur",
+    "temperature": "température",
+    "threshold:音量小于这个值视作静音的备选切割点": "seuil: le volume inférieur à cette valeur est considéré comme un point de coupe silencieux alternatif",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "Le mode n'est pas encore supporté par la v3. Une erreur se produira en cas d'utilisation.",
+    "v3输出如果觉得闷可以试试开超分": "Si la sortie v3 semble étouffée, essayez l'upscaling",
+    "不切": "Pas de découpe",
+    "中文": "Chinois",
+    "中文教程文档": "Documentation du tutoriel en chinois",
+    "中英混合": "Mélange de chinois et d'anglais",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Audio de référence principal (Veuillez télécharger un audio de référence entre 3 et 10 secondes, dépasser cette limite entraînera une erreur !)",
+    "主参考音频的文本": "Texte de l’audio de référence principal",
+    "主参考音频的语种": "Langue de l’audio de référence principal",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "Également possible d'entrer en lot des fichiers audio, au choix, privilégiez la lecture du dossier",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Traitement par lot de séparation voix-accompagnement en utilisant le modèle UVR5.",
+    "人声分离WebUI": "Interface Web de séparation des voix",
+    "人声提取激进程度": "Degré d'extraction des voix",
+    "以下文件或文件夹不存在": "Aucun Fichier ou Dossier de ce Type",
+    "以下模型不存在:": "Aucun Modèle de ce Type:",
+    "伴奏人声分离&去混响&去回声": "Séparation de la voix et de l'accompagnement, suppression de la réverbération et de l'écho",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Vous n'avez pas téléchargé les paramètres du modèle d'upscaling, donc l'upscaling ne sera pas effectué. Veuillez suivre le tutoriel pour télécharger les fichiers nécessaires",
+    "使用无参考文本模式时建议使用微调的GPT": "Il est recommandé d'utiliser un GPT finement ajusté lors de l'utilisation du mode sans texte de référence.",
+    "保存频率save_every_epoch": "Fréquence de sauvegarde (sauvegarder à chaque époque)",
+    "保持随机": "Garder aléatoire",
+    "关闭": "Fermer ",
+    "凑50字一切": "Assembler 50 mots tout",
+    "凑四句一切": "Composez quatre phrases pour tout remplir",
+    "分桶处理模式已关闭": "Mode de traitement par regroupement désactivé",
+    "分桶处理模式已开启": "Mode de traitement par regroupement activé",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "Le mode de retour segmenté ne prend pas en charge le regroupement des données, désactivation automatique",
+    "分段返回模式已开启": "Mode de retour segmenté activé",
+    "分段间隔(秒)": "Intervalle de segmentation (secondes)",
+    "分段间隔过小，已自动设置为0.01": "L’intervalle de segmentation est trop court, réglé automatiquement à 0.01",
+    "切分": "Segmentation",
+    "切分后文本": "Texte après segmentation",
+    "切分后的子音频的输出根目录": "Répertoire racine de sortie des sous-audios après découpage",
+    "切分文本": "Segmenter le texte",
+    "切割使用的进程数": "Nombre de processus utilisés pour le découpage",
+    "刷新模型路径": "Actualiser le chemin du modèle",
+    "前端处理后的文本(每句):": "Texte après traitement frontal (par phrase):",
+    "前置数据集获取工具": "Outil de récupération des ensembles de données",
+    "占用中": " Occupé",
+    "去混响/去延迟，附：": "Suppression de la réverbération / suppression du retard, ci-joint:",
+    "参考音频在3~10秒范围外，请更换！": "Veuillez remplacer l'audio de référence si sa durée est en dehors de la plage de 3 à 10 secondes!",
+    "参考音频的文本": "Texte de l'audio de référence",
+    "参考音频的语种": "Langue de l'audio de référence",
+    "句间停顿秒数": "Temps de pause entre les phrases (secondes)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Optionnel : Téléchargez plusieurs fichiers audio de référence en les faisant glisser (recommandé d'être du même genre) et fusionnez leur tonalité. Si cette option est laissée vide, la tonalité sera contrôlée par l'unique fichier audio de référence à gauche. Si vous ajustez le modèle, il est recommandé que tous les fichiers audio de référence aient des tonalités dans l'ensemble d'entraînement d'ajustement ; le modèle pré-entrainé peut être ignoré.",
+    "合成语音": "Synthèse vocale",
+    "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)。": "Exemple de format de chemin de dossier valide : E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (copiez-le depuis la barre d'adresse de l'explorateur de fichiers).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "Le support pour la conversion phonémique, l’édition manuelle des phonèmes et la synthèse vocale par étapes sera ajouté ultérieurement.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Si vous ne comprenez pas bien l'audio de référence (vous ne savez pas quoi écrire), vous pouvez activer cette option. Une fois activée, le texte de référence sera ignoré.",
+    "启用并行推理版本": "Activer la version d’inférence parallèle",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Veuillez indiquer le répertoire contenant les audio découpés ! Le chemin complet du fichier audio à lire = ce répertoire - nom du fichier correspondant à l'onde dans le fichier .list (pas le chemin complet). Si laissé vide, le chemin absolu dans le fichier .list sera utilisé.",
+    "多语种混合": "Mélange multilingue",
+    "多语种混合(粤语)": "Mélange Multilingue (Cantonais)",
+    "失败": " Échec",
+    "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.": "Si vous n'acceptez pas ces conditions, vous ne pouvez ni utiliser ni citer aucun code ou fichier du package logiciel. Voir LICENSE à la racine.",
+    "实际输入的参考文本:": "Texte de référence réellement saisi:",
+    "实际输入的目标文本(切句后):": "Texte cible réellement saisi (après découpage):",
+    "实际输入的目标文本(每句):": "Texte cible réellement saisi (par phrase):",
+    "实际输入的目标文本:": "Texte cible réellement saisi:",
+    "导出文件格式": "Format d'exportation du fichier",
+    "已关闭": " Fermé",
+    "已完成": " Terminé",
+    "已开启": " Activé",
+    "并行推理": "Inférence parallèle",
+    "并行推理模式已关闭": "Mode d’inférence parallèle désactivé",
+    "并行推理模式已开启": "Mode d’inférence parallèle activé",
+    "开启": "Activer ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Activer le mode sans texte de référence. Laisser le texte de référence vide équivaut également à activer le mode.",
+    "微调训练": "Entraînement de fine-tuning",
+    "怎么切": "Comment découper",
+    "总训练轮数total_epoch": "Nombre total d'époques d'entraînement",
+    "总训练轮数total_epoch，不建议太高": "Nombre total d'époques d'entraînement, pas recommandé d'être trop élevé",
+    "指定输出主人声文件夹": "Spécifier le dossier de sortie pour la voix principale",
+    "指定输出非主人声文件夹": "Spécifier le dossier de sortie pour la non-voix principale",
+    "按中文句号。切": "Couper selon les points en chinois.",
+    "按标点符号切": "Couper selon les signes de ponctuation",
+    "按英文句号.切": "Découpez par des points en anglais",
+    "推理": "Inférence",
+    "推理设置": "Paramètres d’inférence",
+    "提取文本Bert特征": "Extraire les caractéristiques du texte avec BERT",
+    "数据分桶(并行推理时会降低一点计算量)": "Regroupement des données (Réduit le coût de calcul en inférence parallèle)",
+    "数据类型精度": "précision du type de données",
+    "文本分词与特征提取": "Segmentation et extraction de caractéristiques du texte",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Outil de segmentation de texte. Un texte trop long peut donner un mauvais résultat, il est donc recommandé de le segmenter. La synthèse se fera selon les sauts de ligne puis sera assemblée.",
+    "文本模块学习率权重": "Poids du taux d'apprentissage du module de texte",
+    "施工中，请静候佳音": "En construction, veuillez attendre patiemment",
+    "日文": "Japonais",
+    "日英混合": "Mélange Japonais-Anglais",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Faut-il ne conserver que les derniers fichiers de poids pour économiser de l'espace disque ?",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Sauvegarder le petit modèle final dans le dossier weights à chaque point de sauvegarde",
+    "是否开启DPO训练选项(实验性)": "Activer l'option d'entraînement DPO (expérimental) ?",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Ajuster la vitesse de parole et la tonalité du dernier résultat de synthèse pour prévenir l'aléatoire.",
+    "显卡信息": "Informations sur la carte graphique",
+    "未下载模型": "Modèle non téléchargé",
+    "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.": "Ce logiciel est open-source sous licence MIT. L'auteur n'exerce aucun contrôle sur le logiciel. L'utilisateur et toute personne diffusant les sorties audio générées sont entièrement responsables.",
+    "标注文件路径 (含文件后缀 *.list)": "Chemin du fichier d'annotation (avec l'extension *.list)",
+    "模型": "Modèle",
+    "模型分为三类：": "Les modèles sont classés en trois catégories:",
+    "模型切换": "Changement de modèle",
+    "每张显卡的batch_size": "Taille de lot par carte graphique",
+    "版本": "Version",
+    "粤英混合": "Mélange Cantonais-Anglais",
+    "粤语": "Cantonais",
+    "终止合成": "Terminer la synthèse",
+    "缺少Hubert数据集": "Jeu de Données Hubert Manquant",
+    "缺少语义数据集": "Jeu de Données Sémantiques Manquant",
+    "缺少音素数据集": "Jeu de Données de Phonèmes Manquant",
+    "缺少音频数据集": "Jeu de Données Audio Manquant",
+    "英文": "Anglais",
+    "训练集格式化一键三连": "Formatage de l'ensemble d'entraînement en un clic",
+    "训练集格式化工具": "Outil de formatage des ensembles d'entraînement",
+    "语义Token提取": "Extraction de tokens sémantiques",
+    "语速": "Débit de parole",
+    "语速调整，高为更快": "Ajuster la vitesse de parole, plus élevée pour plus rapide",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Le réglage de la vitesse vocale ne prend pas en charge le regroupement des données, désactivation automatique",
+    "语音切分": "Segmentation vocale",
+    "语音切分工具": "Outil de segmentation vocale",
+    "语音文本校对标注工具": "Outil d'annotation et de correction des transcriptions vocales",
+    "语音自监督特征提取": "Extraction de caractéristiques auto-supervisée pour l'audio",
+    "语音识别": "Reconnaissance vocale",
+    "语音识别工具": "Outil de reconnaissance vocale",
+    "语音降噪": "Réduction du bruit audio",
+    "语音降噪工具": "Outil de réduction du bruit audio",
+    "请上传3~10秒内参考音频，超过会报错！": "Veuillez télécharger une référence audio de 3 à 10 secondes ; les fichiers plus longs généreront une erreur!",
+    "请上传参考音频": "Veuillez télécharger l'audio de référence",
+    "请填入推理文本": "Veuillez remplir le texte cible",
+    "请填入正确的List路径": "Veuillez Remplir le Chemin Correct de la Liste",
+    "请填入正确的音频文件夹路径": "Veuillez Remplir le Chemin Correct du Dossier Audio",
+    "请输入有效文本": "Veuillez entrer un texte valide",
+    "路径不存在,使用默认配置": "Chemin introuvable, utilisation de la configuration par défaut",
+    "路径不能为空": "Chemin Non Vide Attendu",
+    "路径错误": "Erreur de Chemin",
+    "转换": "Conversion",
+    "辅参考音频(可选多个，或不选)": "Audio de référence secondaire (Facultatif, plusieurs possibles ou aucun)",
+    "输入待处理音频文件夹路径": "Entrez le chemin du dossier audio à traiter",
+    "输入文件夹路径": "Chemin du dossier à entrer",
+    "输入路径不存在": "Le chemin d'entrée n'existe pas",
+    "输入路径存在但不可用": "Le chemin d'entrée existe mais est inutilisable",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "Les fichiers et dossiers commençant par 23456 devraient être présents dans le répertoire logs/nom de l'expérience",
+    "输出信息": "Sortie d'information",
+    "输出文件夹路径": "Chemin du dossier de sortie",
+    "输出的语音": "Audio de sortie",
+    "运行中": " en cours d'exécution",
+    "进度": "Progression",
+    "进程已终止": " Processus terminé",
+    "进程输出信息": " Sortie du processus",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Choisissez le modèle entraîné stocké sous SoVITS_weights et GPT_weights. Par défaut, l'un d'eux est un modèle de base pour l'expérience de TTS Zero Shot de 5 secondes.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Nombre d’étapes d’échantillonnage : si le son est bruité, essayez d’augmenter, si c’est lent, essayez de réduire",
+    "重复惩罚": "Pénalité de répétition",
+    "随机种子": "Graine aléatoire",
+    "需先终止才能开启下一次任务": "Il faut d'abord arrêter le processus avant de lancer une nouvelle tâche",
+    "需要合成的切分前文本": "Texte avant segmentation pour la synthèse",
+    "需要合成的文本": "Texte à synthétiser",
+    "需要合成的文本的语种": "Langue du texte à synthétiser",
+    "需要合成的语种": "Langue de synthèse requise",
+    "韩文": "Coreano",
+    "韩英混合": "Mezcla Coreano-Inglés",
+    "音频加载失败": "Échec du Chargement de l'Audio",
+    "音频文件不存在，跳过：": "Fichier audio introuvable, passage : ",
+    "音频标注WebUI": "Interface Web d'annotation audio",
+    "音频自动切分输入路径，可文件可文件夹": "Chemin d'entrée automatique de découpage audio, peut être un fichier ou un dossier",
+    "音频超分中": "Upscaling audio en cours",
+    "预训练GPT模型路径": "Chemin du modèle GPT pré-entraîné",
+    "预训练SSL模型路径": "Chemin du modèle SSL pré-entraîné",
+    "预训练SoVITS-D模型路径": "Chemin du modèle SoVITS-D pré-entraîné",
+    "预训练SoVITS-G模型路径": "Chemin du modèle SoVITS-G pré-entraîné",
+    "预训练中文BERT模型路径": "Chemin du modèle BERT chinois pré-entraîné"
+}
diff --git a/tools/i18n/locale/it_IT.json b/tools/i18n/locale/it_IT.json
new file mode 100644
index 0000000000000000000000000000000000000000..864a74a536f452948409d328845e14a2322671e4
--- /dev/null
+++ b/tools/i18n/locale/it_IT.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net (onnx_dereverb): È la scelta migliore per la riverberazione a due canali, ma non può rimuovere la riverberazione a canale singolo;",
+    "(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。": "(234)DeEcho: Rimuove gli effetti di ritardo. Aggressive è più completo di Normal nella rimozione, DeReverb rimuove ulteriormente la riverberazione, può rimuovere la riverberazione a canale singolo, ma non rimuove completamente la riverberazione a piastra ad alta frequenza.",
+    "*实验/模型名": "*Nome dell'esperimento/modello",
+    "*文本标注文件": "*File di annotazione del testo",
+    "*训练集音频文件目录": "*Directory dei file audio del set di addestramento",
+    "*请上传并填写参考信息": "*Carica e compila le informazioni di riferimento",
+    "*请填写需要合成的目标文本和语种模式": "*Si prega di inserire il testo di destinazione da sintetizzare e la modalità lingua",
+    ".限制范围越小判别效果越好。": "Meno multilingue è meglio",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. Il tempo di elaborazione del modello DeEcho-DeReverb è quasi il doppio di quello degli altri due modelli DeEcho;",
+    "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；": "1. Conserva la voce principale: scegli questa opzione per audio senza armonie, poiché conserva meglio la voce principale rispetto al modello HP5. Include due modelli integrati, HP2 e HP3. HP3 potrebbe far passare leggermente l'accompagnamento ma conserva meglio la voce principale rispetto a HP2;",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-Voce modificata",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. Il modello MDX-Net-Dereverb è piuttosto lento;",
+    "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；": "2. Solo conserva la voce principale: scegli questa opzione per audio con armonie, poiché potrebbe indebolire la voce principale. Include un modello HP5;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. La configurazione più pulita consigliata è MDX-Net seguito da DeEcho-Aggressive.",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Modello per rimuovere la riverberazione e il ritardo (by FoxJoy):",
+    "ASR 模型": "Modello ASR",
+    "ASR 模型尺寸": "Dimensioni del modello ASR",
+    "ASR 语言设置": "Impostazioni linguistiche ASR",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "Addestramento GPT: i pesi del modello sono in GPT_weights/",
+    "GPT模型列表": "Elenco dei modelli GPT",
+    "GPT训练": "Addestramento GPT",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "Parametri di campionamento di GPT (non troppo bassi quando non c'è testo di riferimento. Utilizzare i valori predefiniti in caso di incertezza):",
+    "GPU卡号,只能填1个整数": "Numero della scheda grafica, può essere inserito solo un numero intero",
+    "GPU卡号以-分割，每个卡号一个进程": "Numero di GPU separati da '-'; ogni numero corrisponde a un processo",
+    "LoRA秩": "Rango LoRA",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "Modello base SoVITS V3 mancante, impossibile caricare i pesi LoRA corrispondenti",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "Addestramento SoVITS: i pesi del modello sono in SoVITS_weights/",
+    "SoVITS模型列表": "Elenco dei modelli SoVITS",
+    "SoVITS训练": "Addestramento SoVITS",
+    "TTS推理WebUI": "Interfaccia Web per inferenza TTS",
+    "UVR5人声伴奏分离&去混响去延迟工具": "Strumento UVR5 per separazione voce/accompagnamento & rimozione riverbero e latenza",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix: Quanta proporzione dell'audio normalizzato deve essere miscelata",
+    "batch_size": "Dimensione del batch",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: Come calcolare la curva del volume. Più piccolo è, maggiore è la precisione ma aumenta la complessità computazionale (non significa che una maggiore precisione dà risultati migliori)",
+    "max:归一化后最大值多少": "max: Massimo valore dopo la normalizzazione",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept: Massima durata del silenzio dopo il taglio",
+    "min_interval:最短切割间隔": "min_interval: Intervallo minimo di taglio",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: Lunghezza minima per segmento; se il primo segmento è troppo corto, sarà unito ai segmenti successivi fino a superare questo valore",
+    "temperature": "temperatura",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold: Punto di taglio alternativo considerato silenzioso se il volume è inferiore a questo valore",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "V3 non supporta questa modalità, utilizzarla genererà un errore.",
+    "v3输出如果觉得闷可以试试开超分": "Se l'output v3 sembra ovattato, prova ad attivare il super-risoluzione",
+    "不切": "Nessuna suddivisione",
+    "中文": "Cinese",
+    "中文教程文档": "Documentazione del tutorial in cinese",
+    "中英混合": "Cinese e inglese misti",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Audio di riferimento principale (Si prega di caricare un audio di riferimento tra 3 e 10 secondi, superato questo limite si verificherà un errore!)",
+    "主参考音频的文本": "Testo dell'audio di riferimento principale",
+    "主参考音频的语种": "Lingua dell'audio di riferimento principale",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "È possibile anche inserire file audio in batch, una delle due opzioni, con priorità alla lettura della cartella",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Separazione voce-accompagnamento in batch, utilizza il modello UVR5.",
+    "人声分离WebUI": "Interfaccia Web per separazione vocale",
+    "人声提取激进程度": "Grado di aggressività dell'estrazione vocale",
+    "以下文件或文件夹不存在": "Nessun File o Cartella di Questo Tipo",
+    "以下模型不存在:": "Nessun Modello del Genere:",
+    "伴奏人声分离&去混响&去回声": "Separazione tra accompagnamento e voce & Rimozione dell'eco & Rimozione dell'eco",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Non hai scaricato i parametri del modello di super-risoluzione, quindi l'upscaling non verrà eseguito. Segui il tutorial per scaricare i file necessari.",
+    "使用无参考文本模式时建议使用微调的GPT": "Si consiglia di usare un GPT fine-tuned quando si usa la modalità senza testo di riferimento.",
+    "保存频率save_every_epoch": "Frequenza di salvataggio ogni epoca",
+    "保持随机": "Mantieni casuale",
+    "关闭": "Chiudi ",
+    "凑50字一切": "Riempire con 50 caratteri per tutto",
+    "凑四句一切": "Riempire con quattro frasi per tutto",
+    "分桶处理模式已关闭": "Modalità di elaborazione per bucket disabilitata",
+    "分桶处理模式已开启": "Modalità di elaborazione per bucket abilitata",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "La modalità di ritorno segmentato non supporta l'elaborazione per bucket, l'elaborazione per bucket è stata disabilitata automaticamente",
+    "分段返回模式已开启": "Modalità di ritorno segmentato abilitata",
+    "分段间隔(秒)": "Intervallo di segmentazione (secondi)",
+    "分段间隔过小，已自动设置为0.01": "L'intervallo di segmentazione è troppo breve, impostato automaticamente a 0.01",
+    "切分": "Segmentazione",
+    "切分后文本": "Testo dopo segmentazione",
+    "切分后的子音频的输出根目录": "Directory radice di output per gli audio segmentati",
+    "切分文本": "Segmenta il testo",
+    "切割使用的进程数": "Numero di processi utilizzati per il taglio",
+    "刷新模型路径": "Aggiorna il percorso del modello",
+    "前端处理后的文本(每句):": "Testo elaborato dal front-end (per frase):",
+    "前置数据集获取工具": "Strumento per acquisizione dataset",
+    "占用中": " Occupato",
+    "去混响/去延迟，附：": "Rimozione della riverberazione/ritardo, allegato:",
+    "参考音频在3~10秒范围外，请更换！": "L'audio di riferimento è al di fuori dell'intervallo di 3-10 secondi. Si prega di cambiarlo!",
+    "参考音频的文本": "Testo dell'audio di riferimento",
+    "参考音频的语种": "Lingua dell'audio di riferimento",
+    "句间停顿秒数": "Durata pausa tra le frasi (secondi)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Opzionale: Carica più file audio di riferimento trascinandoli (si consiglia dello stesso genere) e media il loro tono. Se questa opzione è lasciata vuota, il tono sarà controllato dal singolo file audio di riferimento a sinistra. Se si sta perfezionando il modello, è consigliato che tutti i file audio di riferimento abbiano toni presenti nel set di addestramento per il perfezionamento; il modello pre-addestrato può essere ignorato.",
+    "合成语音": "Sintesi vocale",
+    "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)。": "Formato di percorso della cartella valido: E:\\codes\\py39\\vits_vc_gpu\\Esempio di test di BaiLuShuangHua (copiare direttamente dalla barra degli indirizzi del gestore file).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "In futuro verrà aggiunto il supporto per la conversione dei fonemi, la modifica manuale dei fonemi e la sintesi vocale passo dopo passo.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Se l'audio di riferimento non è chiaro o non sai cosa scrivere, abilita questa opzione per ignorare il testo di riferimento.",
+    "启用并行推理版本": "Abilita versione di inferenza parallela",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Inserisci la directory dell'audio segmentato! Il percorso completo del file audio letto = questa directory - unione del nome del file corrispondente alle forme d'onda nel file .list (non il percorso completo). Se lasciato vuoto, verrà utilizzato il percorso assoluto nel file .list.",
+    "多语种混合": "Mix multilingue",
+    "多语种混合(粤语)": "Misto Multilingue (Cantonese)",
+    "失败": " Fallito",
+    "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.": "Se non accetti questi termini, non puoi utilizzare né citare alcun codice o file del pacchetto software. Vedi LICENSE nella directory principale.",
+    "实际输入的参考文本:": "Testo di riferimento effettivamente inserito:",
+    "实际输入的目标文本(切句后):": "Testo di destinazione effettivamente inserito (dopo il taglio delle frasi):",
+    "实际输入的目标文本(每句):": "Testo di destinazione effettivamente inserito (per frase):",
+    "实际输入的目标文本:": "Testo di destinazione effettivamente inserito:",
+    "导出文件格式": "Formato di esportazione del file",
+    "已关闭": " Chiuso",
+    "已完成": " Completato",
+    "已开启": " Attivato",
+    "并行推理": "Inferenza parallela",
+    "并行推理模式已关闭": "Modalità di inferenza parallela disabilitata",
+    "并行推理模式已开启": "Modalità di inferenza parallela abilitata",
+    "开启": "Attiva ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Attivare la modalità senza testo di riferimento. Anche se non inserisci un testo di riferimento, la modalità verrà attivata.",
+    "微调训练": "Addestramento fine-tuning",
+    "怎么切": "Come tagliare",
+    "总训练轮数total_epoch": "Numero totale di epoche di addestramento",
+    "总训练轮数total_epoch，不建议太高": "Numero totale di epoche di addestramento, non raccomandato troppo alto",
+    "指定输出主人声文件夹": "Specifica la cartella di output per la voce principale",
+    "指定输出非主人声文件夹": "Specifica la cartella di output per la non voce principale",
+    "按中文句号。切": "Taglia secondo il punto cinese.",
+    "按标点符号切": "Taglia secondo i segni di punteggiatura",
+    "按英文句号.切": "Taglia secondo il punto inglese",
+    "推理": "Inferenza",
+    "推理设置": "Impostazioni di inferenza",
+    "提取文本Bert特征": "Estrai caratteristiche Bert dal testo",
+    "数据分桶(并行推理时会降低一点计算量)": "Raggruppamento dei dati (Riduce il costo computazionale nell'inferenza parallela)",
+    "数据类型精度": "precisione del tipo di dati",
+    "文本分词与特征提取": "Tokenizzazione ed estrazione delle caratteristiche del testo",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Strumento di segmentazione del testo. Testi troppo lunghi potrebbero non dare buoni risultati, quindi si consiglia di segmentarli. La sintesi verrà eseguita separando il testo in base ai ritorni a capo e poi concatenandolo.",
+    "文本模块学习率权重": "Peso del tasso di apprendimento del modulo di testo",
+    "施工中，请静候佳音": "In costruzione, attendi pazientemente le buone notizie",
+    "日文": "Giapponese",
+    "日英混合": "Mix giapponese e inglese",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Salvare solo i file di pesi più recenti per risparmiare spazio su disco?",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Salvare il modello finale più piccolo nella cartella weights ad ogni punto di salvataggio",
+    "是否开启DPO训练选项(实验性)": "Attivare l'opzione di addestramento DPO (sperimentale)?",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Regola la velocità del parlato e il tono dell'ultimo risultato di sintesi per prevenire la casualità.",
+    "显卡信息": "Informazioni sulla scheda grafica",
+    "未下载模型": "Modello non scaricato",
+    "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.": "Questo software è open-source sotto licenza MIT. L'autore non esercita alcun controllo sul software. L'utente e chiunque diffonda gli output audio generati sono pienamente responsabili.",
+    "标注文件路径 (含文件后缀 *.list)": "Percorso del file di annotazione (con estensione *.list)",
+    "模型": "Modello",
+    "模型分为三类：": "I modelli sono divisi in tre categorie:",
+    "模型切换": "Cambio del modello",
+    "每张显卡的batch_size": "Batch size per ogni scheda grafica",
+    "版本": "Versione",
+    "粤英混合": "Misto Cantonese-Inglese",
+    "粤语": "Cantonese",
+    "终止合成": "Termina sintesi",
+    "缺少Hubert数据集": "Dataset di Hubert Mancante",
+    "缺少语义数据集": "Dataset Semantico Mancante",
+    "缺少音素数据集": "Dataset di Fonemi Mancante",
+    "缺少音频数据集": "Dataset Audio Mancante",
+    "英文": "Inglese",
+    "训练集格式化一键三连": "Formattazione del dataset di addestramento in un clic",
+    "训练集格式化工具": "Strumento di formattazione dataset",
+    "语义Token提取": "Estrazione token semantici",
+    "语速": "Velocità della voce",
+    "语速调整，高为更快": "Regolare la velocità della voce, più alta per più veloce",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Il controllo della velocità del discorso non supporta l'elaborazione per bucket, l'elaborazione per bucket è stata disabilitata automaticamente",
+    "语音切分": "Segmentazione vocale",
+    "语音切分工具": "Strumento di segmentazione vocale",
+    "语音文本校对标注工具": "Strumento di annotazione e correzione testi vocali",
+    "语音自监督特征提取": "Estrazione auto-supervisionata delle caratteristiche audio",
+    "语音识别": "Riconoscimento vocale",
+    "语音识别工具": "Strumento di riconoscimento vocale",
+    "语音降噪": "Riduzione del rumore audio",
+    "语音降噪工具": "Strumento di riduzione del rumore audio",
+    "请上传3~10秒内参考音频，超过会报错！": "Carica un audio di riferimento della durata compresa tra 3 e 10 secondi. Superiore a questo, verrà generato un errore!",
+    "请上传参考音频": "Si prega di caricare l'audio di riferimento",
+    "请填入推理文本": "Si prega di inserire il testo di destinazione",
+    "请填入正确的List路径": "Si Prega di Inserire il Percorso Corretto della Lista",
+    "请填入正确的音频文件夹路径": "Si Prega di Inserire il Percorso Corretto della Cartella Audio",
+    "请输入有效文本": "Inserisci un testo valido",
+    "路径不存在,使用默认配置": "Percorso non trovato, utilizzo della configurazione predefinita",
+    "路径不能为空": "Percorso Vuoto Non Consentito",
+    "路径错误": "Errore di Percorso",
+    "转换": "Converti",
+    "辅参考音频(可选多个，或不选)": "Audio di riferimento secondario (Facoltativo, seleziona più o nessuno)",
+    "输入待处理音频文件夹路径": "Inserisci il percorso della cartella dei file audio da elaborare",
+    "输入文件夹路径": "Inserisci il percorso della cartella",
+    "输入路径不存在": "Il percorso di input non esiste",
+    "输入路径存在但不可用": "Il percorso di input esiste ma non è utilizzabile",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "Nella cartella logs/nome dell'esperimento dovrebbero esserci file e cartelle che iniziano con 23456",
+    "输出信息": "Informazioni di output",
+    "输出文件夹路径": "Percorso della cartella di output",
+    "输出的语音": "Audio di output",
+    "运行中": " In esecuzione",
+    "进度": "Avanzamento",
+    "进程已终止": " Processo terminato",
+    "进程输出信息": " Output del processo",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Scegli il modello salvato in SoVITS_weights e GPT_weights dopo l'addestramento. Uno di default è il modello di base, utilizzato per l'esperienza di Zero Shot TTS in 5 secondi.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Passi di campionamento: se sembra rumoroso, prova a aumentarlo, se è lento, prova a diminuirlo",
+    "重复惩罚": "Penalità di ripetizione",
+    "随机种子": "Seme casuale",
+    "需先终止才能开启下一次任务": "Devi prima terminare il processo prima di avviare una nuova attività",
+    "需要合成的切分前文本": "Testo prima della segmentazione per la sintesi",
+    "需要合成的文本": "Testo da sintetizzare",
+    "需要合成的文本的语种": "Lingua del testo da sintetizzare",
+    "需要合成的语种": "Lingua da sintetizzare",
+    "韩文": "Coreano",
+    "韩英混合": "Misto Coreano-Inglese",
+    "音频加载失败": "Caricamento Audio Fallito",
+    "音频文件不存在，跳过：": "File audio non trovato, salto: ",
+    "音频标注WebUI": "Interfaccia Web per annotazione audio",
+    "音频自动切分输入路径，可文件可文件夹": "Percorso di input per la segmentazione automatica dell'audio, può essere un file o una cartella",
+    "音频超分中": "Super-risoluzione audio in corso",
+    "预训练GPT模型路径": "Percorso del modello GPT pre-addestrato",
+    "预训练SSL模型路径": "Percorso del modello SSL pre-addestrato",
+    "预训练SoVITS-D模型路径": "Percorso del modello SoVITS-D pre-addestrato",
+    "预训练SoVITS-G模型路径": "Percorso del modello SoVITS-G pre-addestrato",
+    "预训练中文BERT模型路径": "Percorso del modello BERT cinese pre-addestrato"
+}
diff --git a/tools/i18n/locale/ja_JP.json b/tools/i18n/locale/ja_JP.json
new file mode 100644
index 0000000000000000000000000000000000000000..2ce94b953dd47526eaf77cd8c8f1b142bd05c652
--- /dev/null
+++ b/tools/i18n/locale/ja_JP.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb):二重チャンネルのリバーブに最適な選択ですが、単一チャンネルのリバーブは除去できません；",
+    "(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。": "(234)DeEcho:遅延効果を除去します。AggressiveはNormalよりも徹底的に除去し、DeReverbは追加でリバーブを除去し、モノラルリバーブを除去できますが、高周波数のプレートリバーブは完全には除去できません。",
+    "*实验/模型名": "*実験/モデル名",
+    "*文本标注文件": "*テキスト注釈ファイル",
+    "*训练集音频文件目录": "*トレーニングデータのオーディオファイルディレクトリ",
+    "*请上传并填写参考信息": "*参照情報をアップロードして記入してください",
+    "*请填写需要合成的目标文本和语种模式": "*合成対象テキストと言語モードを入力してください",
+    ".限制范围越小判别效果越好。": "多言語対応を減らした方が良い",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1、DeEcho-DeReverbモデルの処理時間は、他の2つのDeEchoモデルのほぼ2倍です；",
+    "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；": "1、主音を保持: ハーモニーなしの音声にはこのオプションを選択し、HP5よりも主音の保持が優れています。HP2とHP3の2つのモデルが内蔵されており、HP3はわずかに伴奏を漏らす可能性がありますが、HP2よりも主音の保持がわずかに良いです；",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-ボイスチェンジャー",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverbモデルはかなり遅いです；",
+    "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；": "2、主音のみを保持: ハーモニー付きの音声にはこのオプションを選択し、主音が弱くなる可能性があります。HP5モデルが1つ内蔵されています；",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3、最もクリーンな設定は、MDX-Netの後にDeEcho-Aggressiveを使用することをお勧めします。",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3、リバーブ除去と遅延除去モデル（by FoxJoy）：",
+    "ASR 模型": "ASR モデル",
+    "ASR 模型尺寸": "ASRモデルサイズ",
+    "ASR 语言设置": "ASR 言語設定",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT トレーニング: モデルの重みファイルは GPT_weights/ にあります",
+    "GPT模型列表": "GPTモデルリスト",
+    "GPT训练": "GPTトレーニング",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT サンプリングパラメーター（参照テキストがない場合はあまり低くしないでください。わからない場合はデフォルトを使用してください）：",
+    "GPU卡号,只能填1个整数": "GPU番号、1つの整数しか入力できません",
+    "GPU卡号以-分割，每个卡号一个进程": "GPUカード番号はハイフンで区切り、各カード番号ごとに1つのプロセスが実行されます",
+    "LoRA秩": "LoRAランク",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 のベースモデルが不足しているため、対応する LoRA の重みをロードできません",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS トレーニング: モデルの重みファイルは SoVITS_weights/ にあります",
+    "SoVITS模型列表": "SoVITSモデルリスト",
+    "SoVITS训练": "SoVITSトレーニング",
+    "TTS推理WebUI": "TTS推論WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5ボーカルアカンパニメント分離＆リバーブおよびディレイ除去ツール",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix：正規化後のオーディオが入る割合",
+    "batch_size": "バッチサイズ",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size: 音量曲線の計算方法、小さいほど精度が高くなりますが、計算量が増加します（精度が高いほど必ずしも効果が良いわけではありません）",
+    "max:归一化后最大值多少": "max：正規化後の最大値",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept：切り終えた後、最大でどれだけ静かにするか",
+    "min_interval:最短切割间隔": "min_interval：最短カット間隔",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length：各セグメントの最小長さ。最初のセグメントが短すぎる場合、連続して後続のセグメントに接続され、この値を超えるまで続きます。",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "閾値：この値未満の音量は静音と見なされ、代替のカットポイントとして扱われます",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3では現在このモードをサポートしておらず、使用するとエラーが発生します。",
+    "v3输出如果觉得闷可以试试开超分": "v3の出力がこもっていると感じた場合、超解像を試してください",
+    "不切": "切らない",
+    "中文": "中国語",
+    "中文教程文档": "中国語チュートリアルドキュメント",
+    "中英混合": "中英混合",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "プライマリーFERエンス音声（3〜10秒以内の参考音声をアップロードしてください。これを超えるとエラーが発生します！）",
+    "主参考音频的文本": "プライマリーFERエンス音声のテキスト",
+    "主参考音频的语种": "プライマリーFERエンス音声の言語",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "複数のオーディオファイルもインポートできます。フォルダパスが存在する場合、この入力は無視されます。",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "人声と伴奏の分離をバッチ処理で行い、UVR5モデルを使用します。",
+    "人声分离WebUI": "ボーカル分離WebUI",
+    "人声提取激进程度": "人声抽出の積極性",
+    "以下文件或文件夹不存在": "そのようなファイルまたはフォルダは存在しません",
+    "以下模型不存在:": "モデルが存在しません:",
+    "伴奏人声分离&去混响&去回声": "ボーカル/伴奏の分離と残響の除去",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "超解像モデルのパラメータをダウンロードしていないため、超解像は行われません。超解像を行いたい場合は、まずチュートリアルを参照してファイルをダウンロードしてください",
+    "使用无参考文本模式时建议使用微调的GPT": "参照テキストなしモードを使用する際は、微調整されたGPTを推奨めることをお勧びします。",
+    "保存频率save_every_epoch": "保存頻度save_every_epoch",
+    "保持随机": "ランダムを維持",
+    "关闭": "閉じる",
+    "凑50字一切": "50文字ずつカット",
+    "凑四句一切": "4つの文で埋める",
+    "分桶处理模式已关闭": "バケット処理モードを無効化",
+    "分桶处理模式已开启": "バケット処理モードを有効化",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "セグメントされた返却モードはバケット処理をサポートしていないため、自動的にバケット処理を無効化",
+    "分段返回模式已开启": "セグメントされた返却モードを有効化",
+    "分段间隔(秒)": "セグメント間隔（秒）",
+    "分段间隔过小，已自动设置为0.01": "セグメント間隔が短すぎていので、自動的に0.01に設定されました",
+    "切分": "セグメント",
+    "切分后文本": "セグメント後のテキスト",
+    "切分后的子音频的输出根目录": "分割後のサブオーディオの出力ルートディレクトリ",
+    "切分文本": "テキストをセグメント",
+    "切割使用的进程数": "分割に使用されるプロセス数",
+    "刷新模型路径": "モデルのパスを更新",
+    "前端处理后的文本(每句):": "フロントエンド処理後のテキスト（文ごと）:",
+    "前置数据集获取工具": "前処理データセット取得ツール",
+    "占用中": "使用中",
+    "去混响/去延迟，附：": "残響除去/遅延除去、附：",
+    "参考音频在3~10秒范围外，请更换！": "参照音声が3～10秒の範囲外です。別の音声に変更してください！",
+    "参考音频的文本": "参照オーディオのテキスト",
+    "参考音频的语种": "参照オーディオの言語",
+    "句间停顿秒数": "文間のポーズ秒数",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "オプション：複数の参照オーディオファイルをドラッグ＆ドロップしてアップロードし、それらのトーンを平均化します（同性推奨）。このオプションを空白のままにした場合、トーンは左側の単一の参照オーディオによって制御されます。モデルを微調整する場合、すべての参照オーディオファイルが微調整のトレーニングセット内のトーンを持つことをお勧めします。プリトレーニングモデルは無視しても構いません。",
+    "合成语音": "推論を開始",
+    "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)。": "適切なフォルダパスの例： E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华テストサンプル（ファイルマネージャのアドレスバーからコピーしてください）。",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "今後、フォンメ转换、手動フォンメ編集、音声合成のステップバイステップ実行をサポートします。",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "参照音声がはっきり、または何を書くかわからない場合は、このオプションを有効にして参照テキストを無視します。",
+    "启用并行推理版本": "並列推論バージョンを有効化",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "切断後の音声ファイルが格納されているディレクトリを入力してください！読み取り対象の音声ファイルの完全パス = このディレクトリ - 結合 - listファイル内の波形に対応するファイル名（完全パスではありません）。空白の場合、.listファイル内の絶対完全パスを使用します。",
+    "多语种混合": "多言語混合",
+    "多语种混合(粤语)": "多言語混合(粤語)",
+    "失败": "失敗",
+    "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.": "この条件を認めない場合、ソフトウェアパッケージ内の任意のコードやファイルを使用または引用することはできません。詳細はルートディレクトリのLICENSEを参照してください。",
+    "实际输入的参考文本:": "実際に入力された参照テキスト：",
+    "实际输入的目标文本(切句后):": "実際に入力された目標テキスト（文分割後）：",
+    "实际输入的目标文本(每句):": "実際に入力された目標テキスト（文ごと）：",
+    "实际输入的目标文本:": "実際に入力された目標テキスト：",
+    "导出文件格式": "エクスポートファイル形式",
+    "已关闭": "閉じました",
+    "已完成": "完了しました",
+    "已开启": "有効化しました",
+    "并行推理": "並列推論",
+    "并行推理模式已关闭": "並列推論モードを無効化",
+    "并行推理模式已开启": "並列推論モードを有効化",
+    "开启": "有効化",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "参照テキストなしモードを有効にします。参照テキストを入力しない場合も同様に有効になります。",
+    "微调训练": "ファインチューニング",
+    "怎么切": "どうやって切るか",
+    "总训练轮数total_epoch": "総トレーニングエポック数total_epoch",
+    "总训练轮数total_epoch，不建议太高": "総トレーニングエポック数total_epoch、高すぎないようにお勧めします",
+    "指定输出主人声文件夹": "ボーカルの出力フォルダを指定:",
+    "指定输出非主人声文件夹": "伴奏の出力フォルダを指定:",
+    "按中文句号。切": "中国語の句点でカット",
+    "按标点符号切": "句読点で分割",
+    "按英文句号.切": "英文のピリオドで切ってください",
+    "推理": "推論",
+    "推理设置": "推論設定",
+    "提取文本Bert特征": "テキストBERT特徴を抽出",
+    "数据分桶(并行推理时会降低一点计算量)": "データバケット化（並列推論時に少し計算コストを減らす）",
+    "数据类型精度": "データ型の精度",
+    "文本分词与特征提取": "テキスト分割と特徴抽出",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "テキストセグメントツール。非常に長いテキストは合成結果が良くない可能性があるため、長いテキストは事前にセグメントすることをお勧びします。合成は改行に基づいて分かれ、その後に結合されます。",
+    "文本模块学习率权重": "テキストモジュールの学習率の重み",
+    "施工中，请静候佳音": "施工中、お待ちください",
+    "日文": "日本語",
+    "日英混合": "日英混合",
+    "是否仅保存最新的权重文件以节省硬盘空间": "最新の重みファイルのみを保存し、ディスクスペースを節約しますか？",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "各保存時間点で最終的な小さなモデルをweightsフォルダに保存するかどうか",
+    "是否开启DPO训练选项(实验性)": "DPO トレーニングオプションを有効にしますか？（実験的）",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "ランダム性を防ぐために、前回の合成結果のスピーチ速度とトーンを調整します。",
+    "显卡信息": "グラフィックカード情報",
+    "未下载模型": "モデルがダウンロードされていません",
+    "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.": "このソフトウェアはMITライセンスでオープンソース化されており、作者はソフトウェアに対して一切の制御権を持っていません。ソフトウェアを使用する者、ソフトウェアから導出される音声を広める者は、自己責任で行ってください。",
+    "标注文件路径 (含文件后缀 *.list)": "ラベル付けファイルのパス（ファイル拡張子 *.list を含む）",
+    "模型": "モデル",
+    "模型分为三类：": "モデルは3種類に分かれています：",
+    "模型切换": "モデル切り替え",
+    "每张显卡的batch_size": "各グラフィックカードのバッチサイズ",
+    "版本": "バージョン",
+    "粤英混合": "粤英混合",
+    "粤语": "粤語",
+    "终止合成": "合成を終了",
+    "缺少Hubert数据集": "Hubertデータセットが欠落しています",
+    "缺少语义数据集": "セマンティクスデータセットが欠落しています",
+    "缺少音素数据集": "音素データセットが欠落しています",
+    "缺少音频数据集": "オーディオデータセットが欠落しています",
+    "英文": "英語",
+    "训练集格式化一键三连": "トレーニングデータのフォーマットを一括処理",
+    "训练集格式化工具": "トレーニングデータフォーマットツール",
+    "语义Token提取": "セマンティックトークン抽出",
+    "语速": "話速",
+    "语速调整，高为更快": "話速調整、高いほど速く",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "音声調整はバケット処理をサポートしていないため、自動的にバケット処理を無効化",
+    "语音切分": "音声分割",
+    "语音切分工具": "音声分割ツール",
+    "语音文本校对标注工具": "音声テキスト校正・ラベル付けツール",
+    "语音自监督特征提取": "音声自己教師あり特徴抽出",
+    "语音识别": "音声認識",
+    "语音识别工具": "音声認識ツール",
+    "语音降噪": "音声ノイズ除去",
+    "语音降噪工具": "音声ノイズ除去ツール",
+    "请上传3~10秒内参考音频，超过会报错！": "3～10秒以内の参照音声をアップロードしてください。それを超えるとエラーが発生します！",
+    "请上传参考音频": "リファレンスオーディオをアップロードしてください",
+    "请填入推理文本": "ターゲットテキストを入力してください",
+    "请填入正确的List路径": "正しいリストパスを入力してください",
+    "请填入正确的音频文件夹路径": "正しいオーディオフォルダパスを入力してください",
+    "请输入有效文本": "有効なテキストを入力してください",
+    "路径不存在,使用默认配置": "パスが見つからないため、デフォルト設定を使用",
+    "路径不能为空": "空のパスは予期されていません",
+    "路径错误": "パスエラー",
+    "转换": "変換",
+    "辅参考音频(可选多个，或不选)": "二次参考音声（複数可、またはなし）",
+    "输入待处理音频文件夹路径": "処理するオーディオフォルダのパスを入力してください:",
+    "输入文件夹路径": "入力フォルダのパス",
+    "输入路径不存在": "入力パスが存在しません",
+    "输入路径存在但不可用": "入力パスは存在しますが、使用できません",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "logs/実験名ディレクトリには23456で始まるファイルとフォルダが含まれている必要があります",
+    "输出信息": "出力情報",
+    "输出文件夹路径": "出力フォルダのパス",
+    "输出的语音": "推論結果",
+    "运行中": "実行中",
+    "进度": "進捗",
+    "进程已终止": "プロセスが終了しました",
+    "进程输出信息": "プロセスの出力情報",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "SoVITS_weightsおよびGPT_weightsに保存されたモデルを選択します。デフォルトのものはプレトレインであり、ゼロショットTTSを体験できます。",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "サンプリングステップ数：ノイズが強い場合は増やし、遅い場合は減らしてみてください",
+    "重复惩罚": "繰り返しペナルティ",
+    "随机种子": "ランダムシード",
+    "需先终止才能开启下一次任务": "次のタスクを開始する前に、現在のタスクを終了する必要があります",
+    "需要合成的切分前文本": "セグメント前の推論テキスト",
+    "需要合成的文本": "推論テキスト",
+    "需要合成的文本的语种": "推論テキストの言語",
+    "需要合成的语种": "推論の言語",
+    "韩文": "韓国語",
+    "韩英混合": "韓英混合",
+    "音频加载失败": "音声の読み込みに失敗しました",
+    "音频文件不存在，跳过：": "オーディオファイルが見つからない。スキップ：",
+    "音频标注WebUI": "音声ラベリングWebUI",
+    "音频自动切分输入路径，可文件可文件夹": "オーディオの自動分割入力パス、ファイルまたはフォルダを指定できます",
+    "音频超分中": "音声超解像中",
+    "预训练GPT模型路径": "事前にトレーニングされたGPTモデルのパス",
+    "预训练SSL模型路径": "事前にトレーニングされたSSLモデルのパス",
+    "预训练SoVITS-D模型路径": "事前にトレーニングされたSoVITS-Dモデルのパス",
+    "预训练SoVITS-G模型路径": "事前にトレーニングされたSoVITS-Gモデルのパス",
+    "预训练中文BERT模型路径": "事前にトレーニングされた中国語BERTモデルのパス"
+}
diff --git a/tools/i18n/locale/ko_KR.json b/tools/i18n/locale/ko_KR.json
new file mode 100644
index 0000000000000000000000000000000000000000..4cf33a6140c80ce18baee66f184fc2b89855fe25
--- /dev/null
+++ b/tools/i18n/locale/ko_KR.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net (onnx_dereverb): 듀얼 채널 리버브에는 가장 적합하지만, 싱글 채널 리버브는 제거할 수 없습니다",
+    "(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。": "(234)DeEcho:지연 효과를 제거합니다. Aggressive는 Normal보다 더 철저하게 제거하며, DeReverb는 추가로 리버브를 제거하여 단일 채널 리버브를 제거할 수 있지만 고주파 리버브는 완전히 제거하지 못합니다.",
+    "*实验/模型名": "*실험/모델 이름",
+    "*文本标注文件": "*텍스트 주석 파일",
+    "*训练集音频文件目录": "*훈련 세트 오디오 파일 디렉터리",
+    "*请上传并填写参考信息": "*참고 정보를 업로드하고 입력하십시오",
+    "*请填写需要合成的目标文本和语种模式": "*합성할 목표 텍스트와 언어 모드를 입력하세요",
+    ".限制范围越小判别效果越好。": "다언어 지원을 줄이는 것이 더 좋습니다",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. DeEcho-DeReverb 모델의 처리 시간은 다른 두 DeEcho 모델의 거의 두 배입니다;",
+    "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；": "1. 사람 목소리를 유지: 화음이 없는 오디오를 선택하면 HP5보다 사람 목소리를 더 잘 유지할 수 있습니다. 내장된 HP2와 HP3 모델이 있으며, HP3는 화음을 약간 놓칠 수 있지만 HP2보다 사람 목소리를 조금 더 잘 유지합니다;",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-음성 변환",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. MDX-Net-Dereverb 모델은 꽤 느립니다;",
+    "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；": "2. 주 목소리만 유지: 화음이 있는 오디오에 이 모델을 선택하면 주 목소리가 약해질 수 있습니다. 내장된 HP5 모델이 있습니다;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. 개인적으로 가장 깨끗한 설정은 먼저 MDX-Net을 사용하고 그 다음에 DeEcho-Aggressive를 사용하는 것입니다;",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. 잔향 제거 및 지연 제거 모델 (by FoxJoy):",
+    "ASR 模型": "ASR 모델",
+    "ASR 模型尺寸": "ASR 모델 크기",
+    "ASR 语言设置": "ASR 언어 설정",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT 훈련: 모델 가중치 파일은 GPT_weights/에 있습니다",
+    "GPT模型列表": "GPT 모델 목록",
+    "GPT训练": "GPT훈련",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT 샘플링 매개변수 (참조 텍스트가 없을 때 너무 낮게 설정하지 마십시오. 확실하지 않으면 기본값을 사용하십시오):",
+    "GPU卡号,只能填1个整数": "GPU 카드 번호, 1개의 정수만 입력 가능",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU 카드 번호는 -로 구분되며 각 카드 번호에 하나의 프로세스가 있어야 함",
+    "LoRA秩": "LoRA 랭크",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 기본 모델이 없어서 해당 LoRA 가중치를 로드할 수 없습니다",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS 훈련: 모델 가중치 파일은 SoVITS_weights/에 있습니다",
+    "SoVITS模型列表": "SoVITS 모델 목록",
+    "SoVITS训练": "SoVITS훈련",
+    "TTS推理WebUI": "TTS 추론 WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5 보컬 및 반주 분리 & 리버브 제거 및 딜레이 제거 도구",
+    "alpha_mix:混多少比例归一化后音频进来": "알파 믹스: 정규화된 오디오가 들어오는 비율",
+    "batch_size": "배치 크기",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop 크기: 볼륨 곡선을 계산하는 방법. 작을수록 정확도가 높아지지만 계산량이 높아집니다 (정확도가 높다고 효과가 좋아지지 않음)",
+    "max:归一化后最大值多少": "최대 값 (정규화 후)",
+    "max_sil_kept:切完后静音最多留多长": "최대 유지되는 정적 길이 (분리 후)",
+    "min_interval:最短切割间隔": "최소 분리 간격",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:각 부분의 최소 길이, 첫 번째 부분이 너무 짧으면 다음 부분과 계속 연결하여 이 값을 초과할 때까지",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "임계 값: 이 값보다 작은 볼륨은 대체 분리 지점으로 간주됩니다.",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3에서는 이 모드를 지원하지 않으며, 사용 시 오류가 발생합니다.",
+    "v3输出如果觉得闷可以试试开超分": "v3 출력이 답답하게 들리면 슈퍼 레졸루션을 켜보세요",
+    "不切": "자르지 않음",
+    "中文": "중국어",
+    "中文教程文档": "중국어 튜토리얼 문서",
+    "中英混合": "중영 혼합",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "주 참조 오디오 (3~10초 이내의 참조 오디오를 업로드하세요. 초과 시 오류 발생!)",
+    "主参考音频的文本": "주 참조 오디오의 텍스트",
+    "主参考音频的语种": "주 참조 오디오의 언어",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "오디오 파일을 일괄로 입력할 수도 있습니다. 둘 중 하나를 선택하고 폴더를 읽기를 우선합니다.",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "보컬과 반주 분리 배치 처리, UVR5 모델 사용.",
+    "人声分离WebUI": "보컬 분리 WebUI",
+    "人声提取激进程度": "보컬 추출의 공격성",
+    "以下文件或文件夹不存在": "해당 파일이나 폴더가 없습니다",
+    "以下模型不存在:": "해당 모델이 존재하지 않습니다:",
+    "伴奏人声分离&去混响&去回声": "반주 및 보컬 분리 & 리버브 제거 & 에코 제거",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "슈퍼 레졸루션 모델의 파라미터를 다운로드하지 않았으므로 슈퍼 레졸루션을 수행하지 않습니다. 사용하려면 먼저 튜토리얼을 참고하여 파일을 다운로드하세요",
+    "使用无参考文本模式时建议使用微调的GPT": "참고 텍스트 없이 사용할 경우 미세 조정된 GPT 사용을 권장합니다",
+    "保存频率save_every_epoch": "저장 빈도 (각 라운드마다)",
+    "保持随机": "랜덤 유지",
+    "关闭": "닫기",
+    "凑50字一切": "50자를 채우십시오",
+    "凑四句一切": "네 문장의 세트를 완성하세요.",
+    "分桶处理模式已关闭": "버킷 처리 모드 비활성화됨",
+    "分桶处理模式已开启": "버킷 처리 모드 활성화됨",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "분할 반환 모드는 버킷 처리를 지원하지 않아 버킷 처리 자동 비활성화됨",
+    "分段返回模式已开启": "분할 반환 모드 활성화됨",
+    "分段间隔(秒)": "분할 간격(초)",
+    "分段间隔过小，已自动设置为0.01": "분할 간격이 너무 짧아 자동으로 0.01로 설정됨",
+    "切分": "분할",
+    "切分后文本": "분할 후 추론 텍스트",
+    "切分后的子音频的输出根目录": "분리된 하위 오디오의 출력 기본 디렉터리",
+    "切分文本": "텍스트 분할",
+    "切割使用的进程数": "사용되는 프로세스 수로 자르기",
+    "刷新模型路径": "모델 경로 새로 고침",
+    "前端处理后的文本(每句):": "프론트엔드 처리 후 텍스트(문장별):",
+    "前置数据集获取工具": "전처리 데이터셋 획득 도구",
+    "占用中": "사용 중",
+    "去混响/去延迟，附：": "리버브 제거/지연 제거, 부록:",
+    "参考音频在3~10秒范围外，请更换！": "참고 오디오가 3~10초 범위를 벗어났습니다. 다른 것으로 바꾸십시오!",
+    "参考音频的文本": "참고 오디오의 텍스트",
+    "参考音频的语种": "참고 오디오의 언어",
+    "句间停顿秒数": "문장 간 정지 시간 (초)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "선택 사항: 여러 참조 오디오 파일을 드래그하여 업로드하고 (동일한 성별을 권장), 그들의 톤을 평균화합니다. 이 옵션을 비워두면 톤은 왼쪽의 단일 참조 오디오로 제어됩니다. 모델을 미세 조정하는 경우 모든 참조 오디오 파일이 미세 조정 훈련 세트 내의 톤을 가지고 있는 것이 좋으며, 사전 훈련된 모델은 무시할 수 있습니다.",
+    "合成语音": "합성 음성",
+    "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)。": "적절한 폴더 경로 형식 예: E:\\codes\\py39\\vits_vc_gpu\\백로서리 테스트 샘플 (파일 관리자 주소 표시줄에서 복사하면 됩니다).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "향후 음소 변환, 수동 음소 편집, 단계별 음성 합성 지원이 추가될 예정입니다.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "참고 오디오의 내용을 정확히 알아들을 수 없을 경우 이 옵션을 활성화하세요. 활성화하면 입력한 참고 텍스트를 무시합니다.",
+    "启用并行推理版本": "병렬 추론 버전 활성화",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "분리된 오디오가 위치한 디렉터리를 입력하세요! 읽어들인 오디오 파일의 전체 경로 = 이 디렉터리 - list 파일에서 파형에 해당하는 파일명(전체 경로가 아님). 비워 두면 .list 파일의 절대 전체 경로를 사용합니다.",
+    "多语种混合": "다국어 혼합",
+    "多语种混合(粤语)": "다국어 혼합(粤語)",
+    "失败": "실패",
+    "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.": "이 조항에 동의하지 않는다면, 소프트웨어 패키지 내의 코드 및 파일을 사용할 수 없습니다. 자세한 내용은 루트 디렉토리 LICENSE를 참조하세요.",
+    "实际输入的参考文本:": "실제 입력된 참고 텍스트:",
+    "实际输入的目标文本(切句后):": "실제 입력된 목표 텍스트(문장 분리 후):",
+    "实际输入的目标文本(每句):": "실제 입력된 목표 텍스트(문장별):",
+    "实际输入的目标文本:": "실제 입력된 목표 텍스트:",
+    "导出文件格式": "내보내기 파일 형식",
+    "已关闭": "닫힘",
+    "已完成": "완료됨",
+    "已开启": "켜짐",
+    "并行推理": "병렬 추론",
+    "并行推理模式已关闭": "병렬 추론 모드 비활성화됨",
+    "并行推理模式已开启": "병렬 추론 모드 활성화됨",
+    "开启": "켜기",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "참고 텍스트 없이 모드를 활성화합니다. 참고 텍스트를 입력하지 않으면 자동으로 활성화됩니다.",
+    "微调训练": "미세 조정 훈련",
+    "怎么切": "자르기 옵션",
+    "总训练轮数total_epoch": "총 훈련 라운드 수 (total_epoch)",
+    "总训练轮数total_epoch，不建议太高": "총 훈련 라운드 수 (total_epoch), 너무 높지 않게 권장됨",
+    "指定输出主人声文件夹": "지정된 주인 목소리 출력 폴더",
+    "指定输出非主人声文件夹": "지정된 비주인 목소리 출력 폴더",
+    "按中文句号。切": "중국어 문장으로 분리하십시오.",
+    "按标点符号切": "구두점을 기준으로 자르기",
+    "按英文句号.切": "영어 문장으로 분리하기",
+    "推理": "추론",
+    "推理设置": "추론 설정",
+    "提取文本Bert特征": "텍스트 BERT 특징 추출",
+    "数据分桶(并行推理时会降低一点计算量)": "데이터 버킷팅 (병렬 추론 시 계산량 감소)",
+    "数据类型精度": "데이터 유형 정밀도",
+    "文本分词与特征提取": "텍스트 토큰화 및 특성 추출",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "텍스트 분할 도구. 너무 긴 텍스트는 합성 결과가 좋지 않을 수 있으므로 분할을 권장합니다. 합성은 텍스트의 줄바꿈을 기준으로 분할된 후 결합됩니다.",
+    "文本模块学习率权重": "텍스트 모듈 학습률 가중치",
+    "施工中，请静候佳音": "공사 중입니다. 기다려주십시오.",
+    "日文": "일본어",
+    "日英混合": "일본어와 영어 혼합",
+    "是否仅保存最新的权重文件以节省硬盘空间": "디스크 공간을 절약하기 위해 최신 가중치 파일만 저장할지 여부",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "각 저장 시간에 최종 작은 모델을 weights 폴더에 저장할지 여부",
+    "是否开启DPO训练选项(实验性)": "DPO 훈련 옵션 활성화 여부 (실험적 기능)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "랜덤성을 방지하기 위해 마지막 합성 결과의 말하기 속도와 톤을 조정합니다.",
+    "显卡信息": "그래픽 카드 정보",
+    "未下载模型": "모델이 다운로드되지 않음",
+    "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.": "본 소프트웨어는 MIT 라이선스로 오픈소스이며, 개발자는 소프트웨어에 대한 어떠한 통제력도 가지지 않습니다. 사용자는 소프트웨어를 이용하거나 이를 통해 생성된 음성을 배포할 경우 모든 책임을 집니다.",
+    "标注文件路径 (含文件后缀 *.list)": "주석 파일 경로 (*.list 확장자 포함)",
+    "模型": "모델",
+    "模型分为三类：": "모델은 3가지로 나뉩니다：",
+    "模型切换": "모델 전환",
+    "每张显卡的batch_size": "각 그래픽 카드의 배치 크기",
+    "版本": "버전",
+    "粤英混合": "粤영 혼합",
+    "粤语": "粤語",
+    "终止合成": "합성 종료",
+    "缺少Hubert数据集": "Hubert 데이터셋이 없습니다",
+    "缺少语义数据集": "의미론 데이터셋이 없습니다",
+    "缺少音素数据集": "음소 데이터셋이 없습니다",
+    "缺少音频数据集": "오디오 데이터셋이 없습니다",
+    "英文": "영어",
+    "训练集格式化一键三连": "훈련 데이터셋 포맷팅 원클릭 실행",
+    "训练集格式化工具": "훈련 데이터셋 포맷팅 도구",
+    "语义Token提取": "의미론적 토큰 추출",
+    "语速": "언어 속도",
+    "语速调整，高为更快": "언어 속도 조정, 높을수록 빠름",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "음속 조절은 버킷 처리를 지원하지 않아 버킷 처리 자동 비활성화됨",
+    "语音切分": "음성 분할",
+    "语音切分工具": "음성 분할 도구",
+    "语音文本校对标注工具": "음성 텍스트 교정 및 주석 도구",
+    "语音自监督特征提取": "음성 자율 학습 특성 추출",
+    "语音识别": "음성 인식",
+    "语音识别工具": "음성 인식 도구",
+    "语音降噪": "음성 잡음 제거",
+    "语音降噪工具": "음성 잡음 제거 도구",
+    "请上传3~10秒内参考音频，超过会报错！": "3~10초 이내의 참고 오디오를 업로드하십시오. 초과하면 오류가 발생합니다!",
+    "请上传参考音频": "참고 오디오를 업로드하세요",
+    "请填入推理文本": "목표 텍스트를 입력하세요",
+    "请填入正确的List路径": "올바른 리스트 경로를 입력하세요",
+    "请填入正确的音频文件夹路径": "올바른 오디오 폴더 경로를 입력하세요",
+    "请输入有效文本": "유효한 텍스트를 입력하세요",
+    "路径不存在,使用默认配置": "경로가 존재하지 않음, 기본 설정 사용",
+    "路径不能为空": "경로가 비어 있을 수 없습니다",
+    "路径错误": "경로 오류",
+    "转换": "변환",
+    "辅参考音频(可选多个，或不选)": "보조 참조 오디오 (여러 개 선택 가능 또는 선택 안 함)",
+    "输入待处理音频文件夹路径": "처리 대기 중인 오디오 폴더 경로 입력",
+    "输入文件夹路径": "폴더 경로 입력",
+    "输入路径不存在": "입력 경로가 존재하지 않음",
+    "输入路径存在但不可用": "입력 경로가 존재하지만 사용할 수 없음",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "logs/실험 이름 디렉터리에는 23456으로 시작하는 파일과 폴더가 있어야 함",
+    "输出信息": "출력 정보",
+    "输出文件夹路径": "출력 폴더 경로",
+    "输出的语音": "출력 음성",
+    "运行中": "실행 중",
+    "进度": "진행 상태",
+    "进程已终止": "프로세스 종료됨",
+    "进程输出信息": "프로세스 출력 정보",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "SoVITS_weights 및 GPT_weights에 저장된 훈련 완료된 모델 중 선택. 기본적으로 하나는 기본 모델이며 5초 Zero Shot TTS를 체험할 수 있습니다.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "샘플링 스텝: 노이즈가 느껴지면 증가, 느리다면 감소 시도",
+    "重复惩罚": "반복 패널티",
+    "随机种子": "랜덤 시드",
+    "需先终止才能开启下一次任务": "다음 작업을 시작하려면 먼저 종료해야 합니다",
+    "需要合成的切分前文本": "분할 전 추론 텍스트",
+    "需要合成的文本": "합성해야 할 텍스트",
+    "需要合成的文本的语种": "추론 텍스트의 언어",
+    "需要合成的语种": "합성해야 할 언어",
+    "韩文": "한국어",
+    "韩英混合": "한영 혼합",
+    "音频加载失败": "오디오 로드 실패",
+    "音频文件不存在，跳过：": "오디오 파일이 존재하지 않음, 건너뜀: ",
+    "音频标注WebUI": "오디오 주석 WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "오디오 자동 분리 입력 경로, 파일 또는 폴더 가능",
+    "音频超分中": "오디오 슈퍼 레졸루션 중",
+    "预训练GPT模型路径": "사전 훈련된 GPT 모델 경로",
+    "预训练SSL模型路径": "사전 훈련된 SSL 모델 경로",
+    "预训练SoVITS-D模型路径": "사전 훈련된 SoVITS-D 모델 경로",
+    "预训练SoVITS-G模型路径": "사전 훈련된 SoVITS-G 모델 경로",
+    "预训练中文BERT模型路径": "사전 훈련된 중국어 BERT 모델 경로"
+}
diff --git a/tools/i18n/locale/pt_BR.json b/tools/i18n/locale/pt_BR.json
new file mode 100644
index 0000000000000000000000000000000000000000..06fb1d608ef09f4cb902d1b4772e9d72f2176d81
--- /dev/null
+++ b/tools/i18n/locale/pt_BR.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net (onnx_dereverb): É a melhor opção para reverberação de dois canais, mas não pode remover a reverberação de um único canal;",
+    "(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。": "(234)DeEcho:Remove os efeitos de atraso. Aggressive é mais completo que Normal na remoção, DeReverb remove adicionalmente a reverberação, pode remover a reverberação de um canal único, mas não remove completamente a reverberação de placa de alta frequência.",
+    "*实验/模型名": "*Nome do experimento/modelo",
+    "*文本标注文件": "*Arquivo de marcação de texto",
+    "*训练集音频文件目录": "*Diretório de arquivos de áudio do conjunto de treinamento",
+    "*请上传并填写参考信息": "Por favor, faça o upload e preencha as informações de referência",
+    "*请填写需要合成的目标文本和语种模式": "*Por favor, insira o texto alvo a ser sintetizado e o modo de idioma.",
+    ".限制范围越小判别效果越好。": "Menos multilinguismo é melhor",
+    "1-GPT-SoVITS-TTS": "1-GPT-SOVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. O tempo de processamento do modelo DeEcho-DeReverb é quase o dobro dos outros dois modelos DeEcho;",
+    "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；": "1. Manter a voz: selecione isso para áudio sem harmonia, que preserva melhor a voz principal do que o HP5. Inclui dois modelos, HP2 e HP3; o HP3 pode permitir um pequeno vazamento de acompanhamento, mas preserva a voz principal um pouco melhor do que o HP2;",
+    "2-GPT-SoVITS-变声": "2-gpt-sovits-mudança de voz",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. O modelo MDX-Net-Dereverb é bastante lento;",
+    "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；": "2. Manter apenas a voz principal: selecione isso para áudio com harmonia, pode haver uma redução na voz principal. Inclui um modelo HP5;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. A configuração mais limpa recomendada é usar primeiro o MDX-Net e depois o DeEcho-Aggressive.",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Modelo de remoção de reverberação e atraso (por FoxJoy):",
+    "ASR 模型": "Modelo ASR",
+    "ASR 模型尺寸": "Tamanho do modelo ASR",
+    "ASR 语言设置": "Configurações de idioma do ASR",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "Treinamento GPT: O arquivo de pesos do modelo está em GPT_weights/",
+    "GPT模型列表": "Lista de modelos GPT",
+    "GPT训练": "Treinamento GPT",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "Parâmetros de amostragem do GPT (não muito baixos quando não houver texto de referência. Use o padrão se não tiver certeza):",
+    "GPU卡号,只能填1个整数": "Número da placa de vídeo, só é possível preencher com um número inteiro",
+    "GPU卡号以-分割，每个卡号一个进程": "Número da placa de vídeo dividido por-, cada número de placa é um processo",
+    "LoRA秩": "Classificação LoRA",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "O modelo base do SoVITS V3 está ausente, impossibilitando o carregamento dos pesos do LoRA",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "Treinamento SoVITS: O arquivo de pesos do modelo está em SoVITS_weights/",
+    "SoVITS模型列表": "Lista de modelos SoVITS",
+    "SoVITS训练": "Treinamento SoVITS",
+    "TTS推理WebUI": "Inferência TTS WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5 Separação de voz e acompanhamento & remoção de reverberação e atraso",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix: Em que proporção o áudio normalizado é misturado de volta",
+    "batch_size": "Tamanho do Lote",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "HOP_SIZE: Como calcular a curva de volume, quanto menor a precisão, maior a quantidade de cálculos (não significa que quanto maior a precisão, melhor o efeito)",
+    "max:归一化后最大值多少": "MAX: Qual é o valor máximo após a normalização?",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept: Depois de cortar, por quanto tempo no máximo o silêncio é mantido",
+    "min_interval:最短切割间隔": "min_interval: O intervalo de corte mínimo",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: Comprimento mínimo de cada segmento. Se o primeiro segmento for muito curto, ele será unido aos segmentos seguintes até exceder este valor",
+    "temperature": "temperatura",
+    "threshold:音量小于这个值视作静音的备选切割点": "Limiar: O volume menor que este valor é considerado como um ponto de corte mudo alternativo",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "O v3 não suporta este modo no momento. O uso resultará em erro.",
+    "v3输出如果觉得闷可以试试开超分": "Se a saída do v3 parecer abafada, tente ativar a super-resolução",
+    "不切": "Não dividir",
+    "中文": "Chinês",
+    "中文教程文档": "Documentação do tutorial em chinês",
+    "中英混合": "Mistura de Chinês e Inglês",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Áudio de Referência Principal (Por favor, carregue um áudio de referência entre 3-10 segundos, exceder esse limite causará um erro!)",
+    "主参考音频的文本": "Texto do Áudio de Referência Principal",
+    "主参考音频的语种": "Idioma do Áudio de Referência Principal",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "Também é possível inserir arquivos de áudio em lote; escolha uma opção, preferencialmente leia a pasta.",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Processamento em lote de separação de voz e acompanhamento, usando o modelo UVR5.",
+    "人声分离WebUI": "WebUI de separação de voz",
+    "人声提取激进程度": "Grau de agressividade da extração de voz",
+    "以下文件或文件夹不存在": "Nenhum Arquivo ou Pasta Encontrado",
+    "以下模型不存在:": "Nenhum Modelo Tal:",
+    "伴奏人声分离&去混响&去回声": "Separação de acompanhamento e voz & remoção de reverberação & remoção de eco",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Os parâmetros do modelo de super-resolução não foram baixados, então a super-resolução não será realizada. Para utilizá-la, siga o tutorial e baixe os arquivos necessários.",
+    "使用无参考文本模式时建议使用微调的GPT": "Recomenda-se o uso de um modelo GPT ajustado ao utilizar o modo sem texto de referência.",
+    "保存频率save_every_epoch": "Frequência de salvamento save_every_epoch",
+    "保持随机": "Manter Aleatório",
+    "关闭": "Fechar ",
+    "凑50字一切": "Complete com 50 caracteres",
+    "凑四句一切": "Complete com quatro frases",
+    "分桶处理模式已关闭": "Modo de Processamento em Balde Desativado",
+    "分桶处理模式已开启": "Modo de Processamento em Balde Ativado",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "Modo de Retorno Segmentado não suporta Processamento em Balde, Processamento em Balde Desativado Automaticamente",
+    "分段返回模式已开启": "Modo de Retorno Segmentado Ativado",
+    "分段间隔(秒)": "Intervalo de Segmentação (Segundos)",
+    "分段间隔过小，已自动设置为0.01": "Intervalo de Segmentação muito curto, automaticamente definido como 0.01",
+    "切分": "Segmentação",
+    "切分后文本": "Texto de Inferência Após Segmentação",
+    "切分后的子音频的输出根目录": "Diretório raiz de saída do sub-áudio após o corte",
+    "切分文本": "Segmentar Texto",
+    "切割使用的进程数": "Número de processos para corte",
+    "刷新模型路径": "Atualizar caminho do modelo",
+    "前端处理后的文本(每句):": "Texto após processamento front-end (por frase):",
+    "前置数据集获取工具": "Ferramenta de obtenção de conjunto de dados pré-processado",
+    "占用中": " Ocupado",
+    "去混响/去延迟，附：": "Remoção de reverberação/remoção de atraso, anexo:",
+    "参考音频在3~10秒范围外，请更换！": "O áudio de referência está fora do intervalo de 3 a 10 segundos. Por favor, substitua!",
+    "参考音频的文本": "Texto do áudio de referência",
+    "参考音频的语种": "Idioma do áudio de referência",
+    "句间停顿秒数": "Tempo de pausa entre frases (segundos)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Opcional: Faça upload de vários arquivos de áudio de referência arrastando e soltando-os (recomendado que sejam do mesmo gênero) e faça uma média dos seus tons. Se essa opção for deixada em branco, o tom será controlado pelo único áudio de referência à esquerda. Se estiver ajustando o modelo, é recomendado que todos os arquivos de áudio de referência tenham tons dentro do conjunto de treinamento de ajuste; o modelo pré-treinado pode ser ignorado.",
+    "合成语音": "Voz sintetizada",
+    "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)。": "Exemplo de formato de caminho de pasta válido: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (copie do endereço da barra do gerenciador de arquivos).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "Suporte para Conversão de Fonemas, Edição Manual de Fonemas e Síntese de Fase por Fase será adicionado no futuro.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Se não conseguir entender claramente o áudio de referência, ative esta opção. Quando ativada, o texto de referência inserido será ignorado.",
+    "启用并行推理版本": "Ativar Versão de Inferência Paralela",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Preencha o diretório onde os áudios cortados estão localizados! O caminho completo dos arquivos de áudio lidos = este diretório - concatenação com o nome do arquivo de forma correspondente no arquivo .list (não o caminho completo). Se deixar em branco, use o caminho absoluto no arquivo .list.",
+    "多语种混合": "Mistura de múltiplos idiomas",
+    "多语种混合(粤语)": "Mistura Multilíngue (Yue)",
+    "失败": " Falha",
+    "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.": "Se você não concorda com estes termos, não poderá usar ou referenciar nenhum código ou arquivo deste pacote de software. Consulte o arquivo LICENSE no diretório raiz para mais detalhes.",
+    "实际输入的参考文本:": "Texto de referência realmente inserido:",
+    "实际输入的目标文本(切句后):": "Texto alvo realmente inserido (após divisão de frases):",
+    "实际输入的目标文本(每句):": "Texto alvo realmente inserido (por frase):",
+    "实际输入的目标文本:": "Texto alvo realmente inserido:",
+    "导出文件格式": "Formato de arquivo de exportação",
+    "已关闭": " Fechado",
+    "已完成": " Concluído",
+    "已开启": " Ativado",
+    "并行推理": "Inferência Paralela",
+    "并行推理模式已关闭": "Modo de Inferência Paralela Desativado",
+    "并行推理模式已开启": "Modo de Inferência Paralela Ativado",
+    "开启": "Ativar ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Ativar o modo sem texto de referência. Não preencher o texto de referência também equivale a ativar.",
+    "微调训练": "Treinamento de ajuste fino",
+    "怎么切": "Como cortar",
+    "总训练轮数total_epoch": "Total de epoch de treinamento",
+    "总训练轮数total_epoch，不建议太高": "Total de epoch de treinamento, não é recomendável um valor muito alto",
+    "指定输出主人声文件夹": "Especificar a pasta de saída da voz principal",
+    "指定输出非主人声文件夹": "Especificar a pasta de saída da voz secundária",
+    "按中文句号。切": "Dividir por ponto final chinês",
+    "按标点符号切": "Dividir por sinais de pontuação",
+    "按英文句号.切": "Dividir por ponto final em inglês",
+    "推理": "Inferência",
+    "推理设置": "Configurações de Inferência",
+    "提取文本Bert特征": "Extrair Características BERT do Texto",
+    "数据分桶(并行推理时会降低一点计算量)": "Agrupamento de Dados (Reduz o Custo Computacional na Inferência Paralela)",
+    "数据类型精度": "precisão do tipo de dado",
+    "文本分词与特征提取": "Tokenização de texto e extração de características",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Ferramenta de Segmentação de Texto. Textos muito longos podem não resultar em uma boa síntese, então é recomendado segmentar. A síntese será feita com base nas quebras de linha e depois concatenada.",
+    "文本模块学习率权重": "Weight da taxa de aprendizado do módulo de texto",
+    "施工中，请静候佳音": "Em construção, por favor, aguarde por um bom som",
+    "日文": "Japonês",
+    "日英混合": "Mistura de Japonês e Inglês",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Deseja salvar apenas os arquivos de pesos mais recentes para economizar espaço em disco?",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Se deve salvar o modelo pequeno final na pasta Weights em cada ponto de salvamento de tempo",
+    "是否开启DPO训练选项(实验性)": "Ativar a opção de treinamento DPO (experimental)?",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Ajuste a velocidade da fala e o tom do último resultado de síntese para evitar aleatoriedade.",
+    "显卡信息": "Informações da placa de vídeo",
+    "未下载模型": "Modelo não baixado",
+    "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.": "Este software é de código aberto sob a licença MIT, e o autor não tem controle sobre seu uso. O usuário e qualquer pessoa que distribua áudio gerado pelo software são totalmente responsáveis.",
+    "标注文件路径 (含文件后缀 *.list)": "Caminho do arquivo de anotações (com a extensão *.list)",
+    "模型": "Modelo",
+    "模型分为三类：": "Modelos dividem-se em três categorias:",
+    "模型切换": "Troca de modelo",
+    "每张显卡的batch_size": "Tamanho do lote de cada placa de vídeo",
+    "版本": "Versão",
+    "粤英混合": "Mistura Yue-Inglês",
+    "粤语": "Yue",
+    "终止合成": "Terminar Síntese",
+    "缺少Hubert数据集": "Conjunto de Dados Hubert Ausente",
+    "缺少语义数据集": "Conjunto de Dados Semânticos Ausente",
+    "缺少音素数据集": "Conjunto de Dados de Fonemas Ausente",
+    "缺少音频数据集": "Conjunto de Dados de Áudio Ausente",
+    "英文": "Inglês",
+    "训练集格式化一键三连": "Formatação do conjunto de treinamento em um clique",
+    "训练集格式化工具": "Ferramenta de formatação do conjunto de treinamento",
+    "语义Token提取": "Extração de tokens semânticos",
+    "语速": "Velocidade da fala",
+    "语速调整，高为更快": "Ajustar a velocidade da fala, mais alta para mais rápido",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Ajuste de Velocidade de Fala não suporta Processamento em Balde, Processamento em Balde Desativado Automaticamente",
+    "语音切分": "Segmentação de áudio",
+    "语音切分工具": "Ferramenta de segmentação de áudio",
+    "语音文本校对标注工具": "Ferramenta de anotação e revisão de texto de áudio",
+    "语音自监督特征提取": "Extração de características de autoaprendizagem de áudio",
+    "语音识别": "Reconhecimento de voz",
+    "语音识别工具": "Ferramenta de reconhecimento de voz",
+    "语音降噪": "Redução de ruído de áudio",
+    "语音降噪工具": "Ferramenta de redução de ruído de áudio",
+    "请上传3~10秒内参考音频，超过会报错！": "Por favor, faça upload de um áudio de referência com duração entre 3 e 10 segundos. Áudios fora dessa faixa causarão erro!",
+    "请上传参考音频": "Por Favor, Carregue o Áudio de Referência",
+    "请填入推理文本": "Por Favor, Preencha o Texto de Inferência",
+    "请填入正确的List路径": "Por Favor, Insira o Caminho Correto da Lista",
+    "请填入正确的音频文件夹路径": "Por Favor, Insira o Caminho Correto da Pasta de Áudio",
+    "请输入有效文本": "Por favor, insira um texto válido",
+    "路径不存在,使用默认配置": "Caminhos Não Encontrados, Usando Configuração Padrão",
+    "路径不能为空": "Esperado Caminho Não Vazio",
+    "路径错误": "Erro de Caminho",
+    "转换": "Converter",
+    "辅参考音频(可选多个，或不选)": "Áudio de Referência Secundário (Múltiplos Opcionais ou Nenhum)",
+    "输入待处理音频文件夹路径": "Caminho da pasta de arquivos de áudio a ser processados",
+    "输入文件夹路径": "Caminho da pasta de entrada",
+    "输入路径不存在": "O caminho de entrada não existe",
+    "输入路径存在但不可用": "O caminho de entrada existe, mas não está disponível",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "Logs de saída/deve haver arquivos e pastas começando com 23456 no diretório do nome do experimento",
+    "输出信息": "Informações de saída",
+    "输出文件夹路径": "Caminho da pasta de saída",
+    "输出的语音": "Áudio de saída",
+    "运行中": " Em execução",
+    "进度": "Progresso",
+    "进程已终止": " Processo encerrado",
+    "进程输出信息": " Informações de saída do processo",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Selecione os modelos armazenados em Sovits_weights e GPT_WEIGHTS. O padrão é o modelo inferior, experiência para 5 segundos de Zero Shot TTS",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Passos de Amostragem: Se parecer ruidoso, tente aumentar; se parecer lento, tente diminuir",
+    "重复惩罚": "Penalidade de Repetição",
+    "随机种子": "Semente Aleatória",
+    "需先终止才能开启下一次任务": "É necessário encerrar primeiro para iniciar a próxima tarefa",
+    "需要合成的切分前文本": "Texto de Inferência Antes da Segmentação",
+    "需要合成的文本": "Texto a ser sintetizado",
+    "需要合成的文本的语种": "Idioma do Texto de Inferência",
+    "需要合成的语种": "Idioma a ser sintetizado",
+    "韩文": "Coreano",
+    "韩英混合": "Mistura Coreano-Inglês",
+    "音频加载失败": "Falha ao Carregar o Áudio",
+    "音频文件不存在，跳过：": "Arquivo de Áudio Não Encontrado, Pulando: ",
+    "音频标注WebUI": "WebUI de anotação de áudio",
+    "音频自动切分输入路径，可文件可文件夹": "Caminho de entrada automático de corte de áudio, pode ser um arquivo ou uma pasta",
+    "音频超分中": "Super-resolução de áudio em andamento",
+    "预训练GPT模型路径": "Caminho do modelo GPT pré-treinado",
+    "预训练SSL模型路径": "Caminho do modelo SSL pré-treinado",
+    "预训练SoVITS-D模型路径": "Caminho do modelo SoVITS-D pré-treinado",
+    "预训练SoVITS-G模型路径": "Caminho do modelo SoVITS-G pré-treinado",
+    "预训练中文BERT模型路径": "Caminho do modelo BERT chinês pré-treinado"
+}
diff --git a/tools/i18n/locale/ru_RU.json b/tools/i18n/locale/ru_RU.json
new file mode 100644
index 0000000000000000000000000000000000000000..830d03131fb2b99bbbf2bef575d5683de6381f6c
--- /dev/null
+++ b/tools/i18n/locale/ru_RU.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb):Это лучший выбор для реверберации с двумя каналами, но он не может устранить реверберацию с одним каналом;",
+    "(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。": "(234)DeEcho:Устраняет эффект задержки. Aggressive устраняет более тщательно, чем Normal, DeReverb дополнительно устраняет реверберацию, может устранить реверберацию с одного канала, но не полностью устраняет высокочастотную реверберацию.",
+    "*实验/模型名": "*Название эксперимента/модели",
+    "*文本标注文件": "*Файл текстовой аннотации",
+    "*训练集音频文件目录": "*Директория аудиофайлов обучающего набора",
+    "*请上传并填写参考信息": "*Пожалуйста, загрузите и заполните референтные данные",
+    "*请填写需要合成的目标文本和语种模式": "*Пожалуйста, введите целевой текст для синтеза и режим языка",
+    ".限制范围越小判别效果越好。": "Чем меньше языков, тем лучше",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. Время обработки модели DeEcho-DeReverb почти вдвое больше, чем у двух других моделей DeEcho;",
+    "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；": "1. Сохранение голоса: выберите этот для аудио без гармоний, сохранение голоса будет лучше, чем HP5. Встроенные модели HP2 и HP3, HP3 может немного пропускать сопровождение, но сохраняет голос немного лучше, чем HP2;",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-переозвучивание",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. Модель MDX-Net-Dereverb довольно медленная;",
+    "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；": "2. Сохранение только основного голоса: выберите это для аудио с гармониями, может ослабить основной голос. Встроенная модель HP5;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. Лично рекомендованная самая чистая конфигурация — сначала MDX-Net, затем DeEcho-Aggressive.",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Модель удаления реверберации и задержек (от FoxJoy):",
+    "ASR 模型": "Модель ASR",
+    "ASR 模型尺寸": "Размер модели ASR",
+    "ASR 语言设置": "Настройки языка ASR",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "Обучение GPT: файлы весов модели находятся в GPT_weights/",
+    "GPT模型列表": "Список моделей GPT",
+    "GPT训练": "Обучение GPT",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "Параметры выборки GPT (не устанавливайте слишком низкие значения, если нет ссылочного текста. Используйте значения по умолчанию, если не уверены):",
+    "GPU卡号,只能填1个整数": "Номер GPU, можно указать только одно целое число",
+    "GPU卡号以-分割，每个卡号一个进程": "Номера GPU разделяются дефисом, на каждый номер отдельный процесс",
+    "LoRA秩": "Ранг LoRA",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "Отсутствует базовая модель SoVITS V3, невозможно загрузить соответствующие веса LoRA",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "Обучение SoVITS: файлы весов модели находятся в SoVITS_weights/",
+    "SoVITS模型列表": "Список моделей SoVITS",
+    "SoVITS训练": "Обучение SoVITS",
+    "TTS推理WebUI": "TTS WebUI для инференса",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5 Инструмент разделения вокала и сопровождения & удаления реверберации и задержки",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:Какая доля нормализованного аудио смешивается",
+    "batch_size": "размер пакета",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:Как рассчитывается кривая громкости, чем меньше, тем выше точность и больше вычислительная нагрузка (большая точность не всегда означает лучший результат)",
+    "max:归一化后最大值多少": "max:Максимальное значение после нормализации",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:Максимальная длительность тишины после разреза",
+    "min_interval:最短切割间隔": "min_interval:Минимальный интервал разреза",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:Минимальная длина каждого отрезка; если первый отрезок слишком короткий, он будет соединен с последующими до достижения этого значения",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:Значение громкости ниже этого считается тишиной для альтернативной точки разреза",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3 пока не поддерживает этот режим, при использовании возникнет ошибка.",
+    "v3输出如果觉得闷可以试试开超分": "Если вывод v3 кажется глухим, попробуйте включить супер-разрешение",
+    "不切": "Не разрезать",
+    "中文": "Китайский",
+    "中文教程文档": "Документация по китайскому руководству",
+    "中英混合": "Китайский и английский",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Основной опорный аудиофайл (пожалуйста, загрузите опорный аудиофайл длительностью от 3 до 10 секунд, превышение этого лимита вызовет ошибку!)",
+    "主参考音频的文本": "Текст основного опорного аудиофайла",
+    "主参考音频的语种": "Язык основного опорного аудиофайла",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "Можно также импортировать несколько аудиофайлов. Если путь к папке существует, то этот ввод игнорируется.",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Обработка разделения вокала и аккомпанемента пакетно с использованием модели UVR5.",
+    "人声分离WebUI": "WebUI для разделения вокала",
+    "人声提取激进程度": "Степень агрессивности извлечения вокала",
+    "以下文件或文件夹不存在": "Такого файла или папки не существует",
+    "以下模型不存在:": "Этот модель не существует",
+    "伴奏人声分离&去混响&去回声": "Разделение вокала/аккомпанемента и удаление эхо",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Параметры модели супер-разрешения не загружены, поэтому супер-разрешение не будет выполняться. Чтобы использовать его, сначала загрузите необходимые файлы, следуя руководству.",
+    "使用无参考文本模式时建议使用微调的GPT": "При использовании режима без эталонного текста рекомендуется использовать дообученный GPT.",
+    "保存频率save_every_epoch": "Частота сохранения save_every_epoch",
+    "保持随机": "Сохранить случайное значение",
+    "关闭": "Закрыть ",
+    "凑50字一切": "Соберите все в 50 символов",
+    "凑四句一切": "Собрать четыре предложения и разрезать",
+    "分桶处理模式已关闭": "Режим обработки пакетов отключен",
+    "分桶处理模式已开启": "Режим обработки пакетов включен",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "Режим сегментированного возврата не поддерживает обработку пакетов, обработка пакетов отключена автоматически",
+    "分段返回模式已开启": "Режим сегментированного возврата включен",
+    "分段间隔(秒)": "Интервал сегментации (секунды)",
+    "分段间隔过小，已自动设置为0.01": "Интервал сегментации слишком короткий, автоматически установлен на 0.01",
+    "切分": "Сегментация",
+    "切分后文本": "Текст для синтеза после сегментации",
+    "切分后的子音频的输出根目录": "Корневой каталог вывода для подаудио после разделения",
+    "切分文本": "Сегментировать текст",
+    "切割使用的进程数": "Количество процессов, используемых для разрезания",
+    "刷新模型路径": "Обновить путь к модели",
+    "前端处理后的文本(每句):": "Текст после предварительной обработки (каждое предложение):",
+    "前置数据集获取工具": "Инструмент для получения предварительных данных",
+    "占用中": " Занято",
+    "去混响/去延迟，附：": "Удаление реверберации/удаление задержки, примечание:",
+    "参考音频在3~10秒范围外，请更换！": "Референтное аудио вне диапазона 3~10 секунд, пожалуйста, замените!",
+    "参考音频的文本": "Текст референтного аудио",
+    "参考音频的语种": "Язык референтного аудио",
+    "句间停顿秒数": "Время паузы между предложениями (в секундах)",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "Необязательно: загрузите несколько файлов с эталонными аудиозаписями, перетащив их (рекомендуется одного пола), и усредните их тон. Если этот параметр не заполнен, тон будет контролироваться одной эталонной аудиозаписью слева. При тонкой настройке модели рекомендуется, чтобы все эталонные аудиозаписи имели тон в пределах обучающего набора для тонкой настройки; предварительно обученную модель можно игнорировать.",
+    "合成语音": "Синтезированный голос",
+    "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)。": "Пример допустимого формата пути к папке: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (просто скопируйте из адресной строки файлового менеджера).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "В будущем будет добавлена поддержка преобразования в фонемы, ручного редактирования фонемов и пошагового выполнения синтеза речи.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Если невозможно разобрать речь в эталонном аудио (и непонятно, что писать), можно включить эту опцию. При включении вводимый эталонный текст будет игнорироваться.",
+    "启用并行推理版本": "Включить параллельную версию вывода",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Заполните каталог, где находятся аудиофайлы после разрезания! Полный путь к читаемым аудиофайлам = каталог - файл .list, имя файла соответствует волне (не полный путь). Если оставить пустым, будет использоваться абсолютный путь из файла .list.",
+    "多语种混合": "Смешанные языки",
+    "多语种混合(粤语)": "Многоязычная смесь (кантонский)",
+    "失败": " Неудача",
+    "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.": "Если вы не согласны с этими условиями, вы не можете использовать или ссылаться на любой код или файлы из этого программного пакета. Подробности смотрите в LICENSE в корневом каталоге.",
+    "实际输入的参考文本:": "Фактически введенный референсный текст:",
+    "实际输入的目标文本(切句后):": "Фактически введенный целевой текст (после разбиения на предложения):",
+    "实际输入的目标文本(每句):": "Фактически введенный целевой текст (каждое предложение):",
+    "实际输入的目标文本:": "Фактически введенный целевой текст:",
+    "导出文件格式": "Формат выходных файлов",
+    "已关闭": " Закрыто",
+    "已完成": " Завершено",
+    "已开启": " Включено",
+    "并行推理": "Параллельный вывод",
+    "并行推理模式已关闭": "Режим параллельного вывода отключен",
+    "并行推理模式已开启": "Режим параллельного вывода включен",
+    "开启": "Включить ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Включить режим без референтного текста. Не заполняя референтный текст, вы также включаете этот режим.",
+    "微调训练": "Обучение с тонкой настройкой",
+    "怎么切": "Как разрезать",
+    "总训练轮数total_epoch": "Общее количество эпох обучения total_epoch",
+    "总训练轮数total_epoch，不建议太高": "Общее количество эпох обучения total_epoch, не рекомендуется слишком высокое",
+    "指定输出主人声文件夹": "Путь к папке для сохранения вокала:",
+    "指定输出非主人声文件夹": "Путь к папке для сохранения аккомпанемента:",
+    "按中文句号。切": "Разделение по китайским точкам.",
+    "按标点符号切": "Разрезать по пунктуационным знакам",
+    "按英文句号.切": "Разрезать по английской точке.",
+    "推理": "Инференс",
+    "推理设置": "Настройки вывода",
+    "提取文本Bert特征": "Извлечь текстовые признаки BERT",
+    "数据分桶(并行推理时会降低一点计算量)": "Разбиение данных на пакеты (уменьшает вычислительные затраты при параллельном выводе)",
+    "数据类型精度": "точность типа данных",
+    "文本分词与特征提取": "Токенизация текста и извлечение признаков",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Инструмент для сегментации текста. Очень длинный текст может не дать хороших результатов синтеза, поэтому рекомендуется сначала разбить его на сегменты. Синтез будет выполняться на основе разрывов строк, а затем объединяться.",
+    "文本模块学习率权重": "Веса скорости обучения текстового модуля",
+    "施工中，请静候佳音": "В разработке, ожидайте хороших новостей",
+    "日文": "Японский",
+    "日英混合": "Японский и английский",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Сохранить только последние файлы весов для экономии дискового пространства?",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Сохранять финальную версию модели в папке weights на каждом этапе сохранения?",
+    "是否开启DPO训练选项(实验性)": "Включить опцию обучения DPO (экспериментально)?",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Настройте скорость речи и тон последнего результата синтеза, чтобы избежать случайности.",
+    "显卡信息": "Информация о видеокарте",
+    "未下载模型": "Модель не загружена",
+    "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.": "Эта программа распространяется с открытым исходным кодом по лицензии MIT, и автор не несёт ответственности за её использование. Пользователь и распространитель звука, созданного программой, несут полную ответственность.",
+    "标注文件路径 (含文件后缀 *.list)": "Путь к файлу разметки (с расширением *.list)",
+    "模型": "Модели",
+    "模型分为三类：": "Модели делятся на три типа:",
+    "模型切换": "Переключение модели",
+    "每张显卡的batch_size": "Размер пакета для каждой видеокарты",
+    "版本": "Версия",
+    "粤英混合": "Кантоно-английская смесь",
+    "粤语": "Кантонийский",
+    "终止合成": "Завершить синтез",
+    "缺少Hubert数据集": "Отсутствует набор данных Hubert",
+    "缺少语义数据集": "Отсутствует семантический набор данных",
+    "缺少音素数据集": "Отсутствует набор данных фонем",
+    "缺少音频数据集": "Отсутствует набор данных аудио",
+    "英文": "Английский",
+    "训练集格式化一键三连": "Форматирование обучающего набора одним кликом",
+    "训练集格式化工具": "Инструмент форматирования обучающего набора",
+    "语义Token提取": "Извлечение семантических токенов",
+    "语速": "Скорость речи",
+    "语速调整，高为更快": "Регулировка скорости речи, чем выше, тем быстрее",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Регулировка скорости речи не поддерживает обработку пакетов, обработка пакетов отключена автоматически",
+    "语音切分": "Разделение аудиофайла",
+    "语音切分工具": "Инструмент разделения аудиофайлов",
+    "语音文本校对标注工具": "Инструмент аннотирования и корректировки аудиотекста",
+    "语音自监督特征提取": "Извлечение характеристик самообучающейся модели речи",
+    "语音识别": "Распознавание речи",
+    "语音识别工具": "Инструмент распознавания речи",
+    "语音降噪": "Шумоподавление аудиофайлов",
+    "语音降噪工具": "Инструмент шумоподавления аудиофайлов",
+    "请上传3~10秒内参考音频，超过会报错！": "Пожалуйста, загрузите референтное аудио длительностью от 3 до 10 секунд, иначе будет ошибка!",
+    "请上传参考音频": "Пожалуйста, загрузите эталонное аудио",
+    "请填入推理文本": "Пожалуйста, введите целевой текст",
+    "请填入正确的List路径": "Пожалуйста, укажите правильный путь к списку",
+    "请填入正确的音频文件夹路径": "Пожалуйста, укажите правильный путь к папке с аудио",
+    "请输入有效文本": "Введите действительный текст",
+    "路径不存在,使用默认配置": "Путь не найден, используется конфигурация по умолчанию",
+    "路径不能为空": "Ожидается, что путь не будет пустым",
+    "路径错误": "Ошибка пути",
+    "转换": "Преобразовать",
+    "辅参考音频(可选多个，或不选)": "Дополнительный опорный аудиофайл (несколько файлов по желанию или ни одного)",
+    "输入待处理音频文件夹路径": "Путь к папке с аудиофайлами для обработки:",
+    "输入文件夹路径": "Введите путь к папке",
+    "输入路径不存在": "Путь ввода не существует",
+    "输入路径存在但不可用": "Путь ввода существует, но недоступен",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "В директории logs/имя_эксперимента должны быть файлы и папки, начинающиеся с 23456",
+    "输出信息": "Статистика",
+    "输出文件夹路径": "Путь к папке для вывода",
+    "输出的语音": "Выводимый звук",
+    "运行中": " Выполняется",
+    "进度": "Прогресс",
+    "进程已终止": " Процесс завершён",
+    "进程输出信息": " Выходные данные процесса",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Выберите модель, сохраненную в SoVITS_weights и GPT_weights после обучения. По умолчанию используется базовая модель для 5-секундного Zero Shot TTS.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Количество шагов выборки: если кажется, что шумно, попробуйте увеличить, если кажется, что медленно, попробуйте уменьшить",
+    "重复惩罚": "Штраф за повторение",
+    "随机种子": "Случайное начальное значение",
+    "需先终止才能开启下一次任务": "Сначала необходимо завершить текущую задачу, прежде чем запускать новую",
+    "需要合成的切分前文本": "Текст для синтеза до сегментации",
+    "需要合成的文本": "Текст для синтеза",
+    "需要合成的文本的语种": "Язык текста для синтеза",
+    "需要合成的语种": "Язык для синтеза",
+    "韩文": "Корейский",
+    "韩英混合": "Корейско-английская смесь",
+    "音频加载失败": "Не удалось загрузить аудио",
+    "音频文件不存在，跳过：": "Аудиофайл не найден, пропускается: ",
+    "音频标注WebUI": "WebUI для аннотирования аудиофайлов",
+    "音频自动切分输入路径，可文件可文件夹": "Путь ввода для автоматического разделения аудио, может быть файлом или папкой",
+    "音频超分中": "Супер-разрешение аудио в процессе",
+    "预训练GPT模型路径": "Путь к предобученной модели GPT",
+    "预训练SSL模型路径": "Путь к предобученной модели SSL",
+    "预训练SoVITS-D模型路径": "Путь к предобученной модели SoVITS-D",
+    "预训练SoVITS-G模型路径": "Путь к предобученной модели SoVITS-G",
+    "预训练中文BERT模型路径": "Путь к предобученной китайской модели BERT"
+}
diff --git a/tools/i18n/locale/tr_TR.json b/tools/i18n/locale/tr_TR.json
new file mode 100644
index 0000000000000000000000000000000000000000..85365ba255251a011f148ae2bb913d8cebfe3961
--- /dev/null
+++ b/tools/i18n/locale/tr_TR.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb):İki kanallı yankılar için en iyi seçimdir, ancak tek kanallı yankıları ortadan kaldıramaz;",
+    "(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。": "(234)DeEcho:Gecikme etkilerini giderir. Aggressive, Normal'dan daha kapsamlı bir şekilde giderir, DeReverb ek olarak yankıyı giderir, tek kanallı yankıyı giderebilir, ancak yüksek frekanslı plaka yankısını tamamen gideremez.",
+    "*实验/模型名": "*Deney/model adı",
+    "*文本标注文件": "*Metin etiketleme dosyası",
+    "*训练集音频文件目录": "*Eğitim seti ses dosyası dizini",
+    "*请上传并填写参考信息": "*Lütfen referans bilgilerini yükleyin ve doldurun",
+    "*请填写需要合成的目标文本和语种模式": "*Lütfen sentezlenecek hedef metni ve dil modunu giriniz.",
+    ".限制范围越小判别效果越好。": "Daha az çok dilli olmak daha iyidir",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1. DeEcho-DeReverb modelinin işleme süresi, diğer iki DeEcho modelinin neredeyse iki katıdır;",
+    "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；": "1. Ses koruma: Arka vokal içermeyen sesler için bu seçeneği kullanın, ana sesi HP5'ten daha iyi korur. HP2 ve HP3 adlı iki model içerir; HP3, arka vokali biraz kaçırabilir ancak ana sesi HP2'ye göre biraz daha iyi korur;",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-Ses Değiştirme",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2. MDX-Net-Dereverb modeli oldukça yavaştır;",
+    "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；": "2. Sadece ana sesi koruma: Arka vokalleri içeren sesler için bu seçeneği kullanın, ana sesi zayıflatabilir. İçinde HP5 modeli var;",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3. Kişisel olarak en temiz konfigürasyon MDX-Net'in ardından DeEcho-Aggressive'dir.",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3. Yankı ve gecikme giderme modeli (FoxJoy tarafından):",
+    "ASR 模型": "ASR modeli",
+    "ASR 模型尺寸": "ASR model boyutu",
+    "ASR 语言设置": "ASR dil ayarları",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT Eğitimi: Model ağırlık dosyaları GPT_weights/ içinde",
+    "GPT模型列表": "GPT model listesi",
+    "GPT训练": "GPT Eğitimi",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT örnekleme parametreleri (referans metin olmadığında çok düşük olmamalıdır. Emin değilseniz varsayılanı kullanın):",
+    "GPU卡号,只能填1个整数": "GPU kart numarası, sadece bir tamsayı girilebilir",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU kart numaraları - ile ayrılır, her kart numarası için bir işlem",
+    "LoRA秩": "LoRA Derecesi",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 temel modeli eksik, ilgili LoRA ağırlıkları yüklenemiyor",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS Eğitimi: Model ağırlık dosyaları SoVITS_weights/ içinde",
+    "SoVITS模型列表": "SoVITS model listesi",
+    "SoVITS训练": "SoVITS Eğitimi",
+    "TTS推理WebUI": "TTS Çıkarım WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5 İnsan Sesli ve Enstrümantal Ayrım & Reverb ve Gecikme Giderme Aracı",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:Normalizasyondan sonraki sesin ne kadarlık bir oranı karıştırılsın",
+    "batch_size": "Toplu Boyut",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:Ses seviyesi eğrisi nasıl hesaplanır, ne kadar küçükse hassasiyet o kadar yüksek ve hesaplama yükü o kadar artar (hassasiyet arttıkça etki mutlaka daha iyi olmaz)",
+    "max:归一化后最大值多少": "max:Normalizasyondan sonra maksimum değer ne kadar",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:Kesimden sonra en fazla ne kadar sessizlik bırakılır",
+    "min_interval:最短切割间隔": "min_interval:Minimum kesim aralığı",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length: bölümün minimum uzunluğu, ilk bölüm çok kısa ise, bu değeri aşana kadar sonraki bölümlerle birleştirilir",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:Ses bu değerden düşükse sessiz olarak kabul edilen alternatif kesim noktası",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3 bu modu desteklemiyor, kullanıldığında hata verecektir.",
+    "v3输出如果觉得闷可以试试开超分": "v3 çıkışı eğer sıkıcıysa, süper çözünürlük açmayı deneyebilirsiniz",
+    "不切": "Kesme",
+    "中文": "Çince",
+    "中文教程文档": "Çince Eğitim Belgesi",
+    "中英混合": "Çince ve İngilizce karışık",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "Ana Referans Sesi (Lütfen 3-10 saniye arasında referans sesi yükleyin, bu sınırı aşarsanız hata alırsınız!)",
+    "主参考音频的文本": "Ana Referans Sesinin Metni",
+    "主参考音频的语种": "Ana Referans Sesinin Dili",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "Ses dosyaları ayrıca toplu olarak, iki seçimle, öncelikli okuma klasörüyle içe aktarılabilir",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "Vokal ve akor ayırma toplu işleme, UVR5 modelini kullanarak.",
+    "人声分离WebUI": "İnsan Sesi Ayrım WebUI",
+    "人声提取激进程度": "Vokal çıkarma agresiflik derecesi",
+    "以下文件或文件夹不存在": "Böyle Bir Dosya veya Klasör Yok",
+    "以下模型不存在:": "Böyle bir model yok:",
+    "伴奏人声分离&去混响&去回声": "Vokal/Müzik Ayrıştırma ve Yankı Giderme",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "Süper çözünürlük model parametrelerini indirmediniz, bu yüzden süper çözünürlük yapılmaz. Süper çözünürlük yapmak için önce kılavuzu takip ederek dosyaları indirin.",
+    "使用无参考文本模式时建议使用微调的GPT": "Referans metinsiz modda, ince ayar yapılmış GPT kullanmanız tavsiye edilir.",
+    "保存频率save_every_epoch": "Kayıt sıklığı save_every_epoch",
+    "保持随机": "Rastgeleliği Koru",
+    "关闭": "Kapat ",
+    "凑50字一切": "50 kelime birleştir ve kes",
+    "凑四句一切": "Dört cümleyi bir araya getirip kes",
+    "分桶处理模式已关闭": "Kovaya İşleme Modu Kapalı",
+    "分桶处理模式已开启": "Kovaya İşleme Modu Etkin",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "Bölümlü Dönüş Modu Kovaya İşlemeyi Desteklemediği İçin Kovaya İşleme Otomatik Olarak Kapatılmıştır",
+    "分段返回模式已开启": "Bölümlü Dönüş Modu Etkin",
+    "分段间隔(秒)": "Bölüm Aralığı (Saniye)",
+    "分段间隔过小，已自动设置为0.01": "Bölüm Aralığı Çok Kısa, Otomatik Olarak 0.01 Olarak Ayarlanmıştır",
+    "切分": "Bölümleme",
+    "切分后文本": "Bölümlenmiş Metin",
+    "切分后的子音频的输出根目录": "Bölündükten sonra alt ses dosyalarının çıktı kök dizini",
+    "切分文本": "Metni Bölümle",
+    "切割使用的进程数": "Kesim için kullanılan işlem sayısı",
+    "刷新模型路径": "Model yolu yenile",
+    "前端处理后的文本(每句):": "Ön işleme tabi tutulan metin (her cümle):",
+    "前置数据集获取工具": "Ön Veri Kümesi Alma Aracı",
+    "占用中": " Kullanımda",
+    "去混响/去延迟，附：": "Yankı giderme/Geçikme giderme, ek:",
+    "参考音频在3~10秒范围外，请更换！": "Referans ses dosyası 3~10 saniye aralığının dışında, lütfen değiştirin!",
+    "参考音频的文本": "Referans ses dosyasının metni",
+    "参考音频的语种": "Referans ses dosyasının dili",
+    "句间停顿秒数": "Cümleler Arası Duraklama Süresi",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "İsteğe bağlı: Birden fazla referans ses dosyasını sürükleyip bırakarak yükleyin (aynı cinsiyetten olmaları önerilir) ve tonlarını ortalayın. Bu seçenek boş bırakılırsa, ton soldaki tek referans ses dosyası tarafından kontrol edilir. Modeli ince ayar yapıyorsanız, tüm referans ses dosyalarının ince ayar eğitim seti içindeki tonlara sahip olması önerilir; önceden eğitilmiş model dikkate alınmayabilir.",
+    "合成语音": "Ses sentezi",
+    "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)。": "Geçerli klasör yolu formatı örneği: E:\\codes\\py39\\vits_vc_gpu\\白鹭霜华测试样例 (dosya yöneticisi adres çubuğundan kopyalayabilirsiniz).",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "Sonraki sürümlerde fonem dönüşümü, el ile fonem düzenleme ve adım adım konuşma sentezi desteği eklenecek.",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "Referans ses kaydını anlamıyorsanız (ne yazacağınızı bilmiyorsanız) açabilirsiniz. Açıldığında yazılmış olan referans metni göz ardı edilir.",
+    "启用并行推理版本": "Paralel Çıkarım Versiyonunu Etkinleştir",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "Kesmeye uygun ses dosyalarının bulunduğu dizini doldurun! Okunan ses dosyasının tam yolu = bu dizin + list dosyasındaki dalga biçimiyle eşleşen dosya adı (tam yol değil). Boş bırakılırsa, .list dosyasındaki tam yol kullanılır.",
+    "多语种混合": "Çok dilli karışım",
+    "多语种混合(粤语)": "Çok dilli karışık (Yue)",
+    "失败": " Başarısız",
+    "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.": "Bu şartları kabul etmiyorsanız, yazılım paketindeki herhangi bir kodu veya dosyayı kullanamaz veya alıntılayamazsınız. Detaylar için ana dizindeki LICENSE dosyasına bakın.",
+    "实际输入的参考文本:": "Gerçekten girilen referans metin:",
+    "实际输入的目标文本(切句后):": "Gerçekten girilen hedef metin (cümleler kesildikten sonra):",
+    "实际输入的目标文本(每句):": "Gerçekten girilen hedef metin (her cümle):",
+    "实际输入的目标文本:": "Gerçekten girilen hedef metin:",
+    "导出文件格式": "Dışa aktarma dosya formatı",
+    "已关闭": " Kapalı",
+    "已完成": " Tamamlandı",
+    "已开启": " Açık",
+    "并行推理": "Paralel Çıkarım",
+    "并行推理模式已关闭": "Paralel Çıkarım Modu Kapalı",
+    "并行推理模式已开启": "Paralel Çıkarım Modu Etkin",
+    "开启": "Aç ",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "Referans metni olmayan mod açık. Referans metni doldurulmazsa bu mod otomatik olarak açılır.",
+    "微调训练": "İnce Ayar Eğitimi",
+    "怎么切": "Nasıl kesilir",
+    "总训练轮数total_epoch": "Toplam eğitim turu sayısı total_epoch",
+    "总训练轮数total_epoch，不建议太高": "Toplam eğitim turu sayısı total_epoch, çok yüksek önerilmez",
+    "指定输出主人声文件夹": "Vokal için çıkış klasörünü belirtin:",
+    "指定输出非主人声文件夹": "Müzik ve diğer sesler için çıkış klasörünü belirtin:",
+    "按中文句号。切": "Çince dönem işaretine göre kes",
+    "按标点符号切": "Noktalama işaretlerine göre kes",
+    "按英文句号.切": "İngilizce nokta işaretine göre kes",
+    "推理": "Çıkarım",
+    "推理设置": "Çıkarım Ayarları",
+    "提取文本Bert特征": "Metin BERT Özelliklerini Çıkar",
+    "数据分桶(并行推理时会降低一点计算量)": "Veri Kovaya Ayrılması (Paralel Çıkarımda Hesaplama Maliyetini Azaltır)",
+    "数据类型精度": "veri türü doğruluğu",
+    "文本分词与特征提取": "Metin Kelimelemeleri ve Özellik Çıkartma",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "Metin Bölümleme Aracı. Çok uzun metinlerin sentez sonuçları iyi olmayabilir, bu nedenle uzun metinleri önce bölümlere ayırmanız önerilir. Sentez, metnin satır sonlarına göre ayrı ayrı yapılıp sonra birleştirilecektir.",
+    "文本模块学习率权重": "Metin modülü öğrenme oranı ağırlığı",
+    "施工中，请静候佳音": "Yapım aşamasında, lütfen iyi haberler için bekleyin",
+    "日文": "Japonca",
+    "日英混合": "Japonca ve İngilizce karışımı",
+    "是否仅保存最新的权重文件以节省硬盘空间": "Sadece en son ağırlık dosyasını kaydedip sabit disk alanı tasarrufu sağlamak isterseniz",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "Her kayıt zamanında son küçük modelin weights klasörüne kaydedilmesi gerekiyor mu",
+    "是否开启DPO训练选项(实验性)": "DPO Eğitim Seçeneğini Açmak (Deneysel)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "Rastgeleliği önlemek için son sentez sonucunun konuşma hızını ve tonunu ayarlayın.",
+    "显卡信息": "Ekran kartı bilgisi",
+    "未下载模型": "Model İndirilmedi",
+    "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.": "Bu yazılım MIT lisansı ile açık kaynak olarak sunulmuştur, yazar yazılım üzerinde herhangi bir kontrol sahibi değildir, yazılımı kullanan veya yazılımın çıktısını dağıtan kişiler tüm sorumluluğu üstlenir.",
+    "标注文件路径 (含文件后缀 *.list)": "Etiketleme Dosyası Yolu (dosya uzantısı *.list dahil)",
+    "模型": "Model",
+    "模型分为三类：": "Modeller üç türdedir:",
+    "模型切换": "Model değiştirme",
+    "每张显卡的batch_size": "Her bir ekran kartı için batch_size",
+    "版本": "Versiyon",
+    "粤英混合": "Yue-İngilizce Karışık",
+    "粤语": "Yue",
+    "终止合成": "Sentezi Sonlandır",
+    "缺少Hubert数据集": "Hubert Veri Seti Eksik",
+    "缺少语义数据集": "Anlamsal Veri Seti Eksik",
+    "缺少音素数据集": "Fonem Veri Seti Eksik",
+    "缺少音频数据集": "Ses Veri Seti Eksik",
+    "英文": "İngilizce",
+    "训练集格式化一键三连": "Eğitim Seti Formatlama Tek Tıkla Üçlü",
+    "训练集格式化工具": "Eğitim Seti Formatlama Aracı",
+    "语义Token提取": "Anlamlı Token Çıkartma",
+    "语速": "Konuşma hızı",
+    "语速调整，高为更快": "Konuşma hızını ayarla, yüksek daha hızlı",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "Konuşma Hızı Ayarlaması Kovaya İşlemeyi Desteklemediği İçin Kovaya İşleme Otomatik Olarak Kapatılmıştır",
+    "语音切分": "Ses Bölme",
+    "语音切分工具": "Ses Bölme Aracı",
+    "语音文本校对标注工具": "Ses Metni Düzeltme Etiketleme Aracı",
+    "语音自监督特征提取": "Ses Kendiliğinden Denetimli Özellik Çıkartma",
+    "语音识别": "Ses Tanıma",
+    "语音识别工具": "Ses Tanıma Aracı",
+    "语音降噪": "Ses Gürültü Azaltma",
+    "语音降噪工具": "Ses Gürültü Azaltma Aracı",
+    "请上传3~10秒内参考音频，超过会报错！": "Lütfen 3~10 saniye arasında bir referans ses dosyası yükleyin, aşım durumunda hata verilecektir!",
+    "请上传参考音频": "Lütfen Referans Sesi Yükleyin",
+    "请填入推理文本": "Lütfen Hedef Metni Girin",
+    "请填入正确的List路径": "Lütfen Doğru Liste Yolunu Girin",
+    "请填入正确的音频文件夹路径": "Lütfen Doğru Ses Klasörü Yolunu Girin",
+    "请输入有效文本": "Geçerli metin girin",
+    "路径不存在,使用默认配置": "Yol Bulunamadı, Varsayılan Yapılandırma Kullanılıyor",
+    "路径不能为空": "Boş Yol Beklenmiyor",
+    "路径错误": "Yol Hatası",
+    "转换": "Dönüştür",
+    "辅参考音频(可选多个，或不选)": "Yardımcı Referans Sesi (İsteğe bağlı birden çok seçilebilir veya hiç seçilmeyebilir)",
+    "输入待处理音频文件夹路径": "İşlenecek ses klasörünün yolunu girin:",
+    "输入文件夹路径": "Dosya klasörü yolu girin",
+    "输入路径不存在": "Girdi Yolu Mevcut Değil",
+    "输入路径存在但不可用": "Girdi Yolu Mevcut Ama Kullanılamaz",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "Çıktı logs/deney adı dizininde 23456 ile başlayan dosya ve klasörler olmalı",
+    "输出信息": "Çıkış bilgisi",
+    "输出文件夹路径": "Çıktı klasörü yolu",
+    "输出的语音": "Çıktı sesi",
+    "运行中": " Çalışıyor",
+    "进度": "İlerleme",
+    "进程已终止": " İşlem Sonlandırıldı",
+    "进程输出信息": " İşlem Çıktı Bilgisi",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "Eğitimi tamamlanmış ve SoVITS_weights ile GPT_weights altına kaydedilmiş modeli seçin. Varsayılan bir temel modeldir, 5 saniyelik Zero Shot TTS deneyimi için kullanılır.",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "Örnekleme Adımları: Eğer gürültülü görünüyorsa, adımları artırın; eğer yavaş görünüyorsa, adımları azaltın",
+    "重复惩罚": "Tekrarlama Cezası",
+    "随机种子": "Rastgele Tohum",
+    "需先终止才能开启下一次任务": "Bir sonraki görevi başlatmadan önce mevcut işlem sonlandırılmalıdır",
+    "需要合成的切分前文本": "Sentez İçin Bölümlenmemiş Metin",
+    "需要合成的文本": "Sentezlenmesi gereken metin",
+    "需要合成的文本的语种": "Sentez Metninin Dili",
+    "需要合成的语种": "Sentezlenmesi gereken dil",
+    "韩文": "Korece",
+    "韩英混合": "Korece-İngilizce Karışık",
+    "音频加载失败": "Ses Yüklenemedi",
+    "音频文件不存在，跳过：": "Ses Dosyası Bulunamadı, Atlanıyor: ",
+    "音频标注WebUI": "Ses Etiketleme WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "Ses otomatik bölme giriş yolu, dosya veya klasör olabilir",
+    "音频超分中": "Ses Süper Çözünürlük Yapılıyor",
+    "预训练GPT模型路径": "Önceden Eğitilmiş GPT Modeli Yolu",
+    "预训练SSL模型路径": "Önceden Eğitilmiş SSL Modeli Yolu",
+    "预训练SoVITS-D模型路径": "Önceden Eğitilmiş SoVITS-D Modeli Yolu",
+    "预训练SoVITS-G模型路径": "Önceden Eğitilmiş SoVITS-G Modeli Yolu",
+    "预训练中文BERT模型路径": "Önceden Eğitilmiş Çince BERT Modeli Yolu"
+}
diff --git a/tools/i18n/locale/zh_CN.json b/tools/i18n/locale/zh_CN.json
new file mode 100644
index 0000000000000000000000000000000000000000..65e99de7e6a76903e7a0dc8b5e509cba24038b42
--- /dev/null
+++ b/tools/i18n/locale/zh_CN.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(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。": "(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。",
+    "*实验/模型名": "*实验/模型名",
+    "*文本标注文件": "*文本标注文件",
+    "*训练集音频文件目录": "*训练集音频文件目录",
+    "*请上传并填写参考信息": "*请上传并填写参考信息",
+    "*请填写需要合成的目标文本和语种模式": "*请填写需要合成的目标文本和语种模式",
+    ".限制范围越小判别效果越好。": ".限制范围越小判别效果越好。",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；",
+    "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；": "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-GPT-SoVITS-变声": "2-GPT-SoVITS-变声",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverb模型挺慢的；",
+    "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；": "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、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3、去混响、去延迟模型（by FoxJoy）：",
+    "ASR 模型": "ASR 模型",
+    "ASR 模型尺寸": "ASR 模型尺寸",
+    "ASR 语言设置": "ASR 语言设置",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT 训练: 模型权重文件在 GPT_weights/",
+    "GPT模型列表": "GPT模型列表",
+    "GPT训练": "GPT训练",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT采样参数(无参考文本时不要太低。不懂就用默认)：",
+    "GPU卡号,只能填1个整数": "GPU卡号,只能填1个整数",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU卡号以-分割，每个卡号一个进程",
+    "LoRA秩": "LoRA秩",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 底模缺失，无法加载相应 LoRA 权重",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS 训练: 模型权重文件在 SoVITS_weights/",
+    "SoVITS模型列表": "SoVITS模型列表",
+    "SoVITS训练": "SoVITS训练",
+    "TTS推理WebUI": "TTS推理WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5人声伴奏分离&去混响去延迟工具",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:混多少比例归一化后音频进来",
+    "batch_size": "batch_size",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）",
+    "max:归一化后最大值多少": "max:归一化后最大值多少",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:切完后静音最多留多长",
+    "min_interval:最短切割间隔": "min_interval:最短切割间隔",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:音量小于这个值视作静音的备选切割点",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3暂不支持该模式，使用了会报错。",
+    "v3输出如果觉得闷可以试试开超分": "v3输出如果觉得闷可以试试开超分",
+    "不切": "不切",
+    "中文": "中文",
+    "中文教程文档": "中文教程文档",
+    "中英混合": "中英混合",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "主参考音频(请上传3~10秒内参考音频，超过会报错！)",
+    "主参考音频的文本": "主参考音频的文本",
+    "主参考音频的语种": "主参考音频的语种",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "也可批量输入音频文件, 二选一, 优先读文件夹",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "人声伴奏分离批量处理， 使用UVR5模型。",
+    "人声分离WebUI": "人声分离WebUI",
+    "人声提取激进程度": "人声提取激进程度",
+    "以下文件或文件夹不存在": "以下文件或文件夹不存在",
+    "以下模型不存在:": "以下模型不存在:",
+    "伴奏人声分离&去混响&去回声": "伴奏人声分离&去混响&去回声",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好",
+    "使用无参考文本模式时建议使用微调的GPT": "使用无参考文本模式时建议使用微调的GPT",
+    "保存频率save_every_epoch": "保存频率save_every_epoch",
+    "保持随机": "保持随机",
+    "关闭": "关闭",
+    "凑50字一切": "凑50字一切",
+    "凑四句一切": "凑四句一切",
+    "分桶处理模式已关闭": "分桶处理模式已关闭",
+    "分桶处理模式已开启": "分桶处理模式已开启",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "分段返回模式不支持分桶处理，已自动关闭分桶处理",
+    "分段返回模式已开启": "分段返回模式已开启",
+    "分段间隔(秒)": "分段间隔(秒)",
+    "分段间隔过小，已自动设置为0.01": "分段间隔过小，已自动设置为0.01",
+    "切分": "切分",
+    "切分后文本": "切分后文本",
+    "切分后的子音频的输出根目录": "切分后的子音频的输出根目录",
+    "切分文本": "切分文本",
+    "切割使用的进程数": "切割使用的进程数",
+    "刷新模型路径": "刷新模型路径",
+    "前端处理后的文本(每句):": "前端处理后的文本(每句):",
+    "前置数据集获取工具": "前置数据集获取工具",
+    "占用中": "占用中",
+    "去混响/去延迟，附：": "去混响/去延迟，附：",
+    "参考音频在3~10秒范围外，请更换！": "参考音频在3~10秒范围外，请更换！",
+    "参考音频的文本": "参考音频的文本",
+    "参考音频的语种": "参考音频的语种",
+    "句间停顿秒数": "句间停顿秒数",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。",
+    "合成语音": "合成语音",
+    "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)。": "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)。",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "后续将支持转音素、手工修改音素、语音合成分步执行。",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。",
+    "启用并行推理版本": "启用并行推理版本",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。",
+    "多语种混合": "多语种混合",
+    "多语种混合(粤语)": "多语种混合(粤语)",
+    "失败": "失败",
+    "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.": "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.",
+    "实际输入的参考文本:": "实际输入的参考文本:",
+    "实际输入的目标文本(切句后):": "实际输入的目标文本(切句后):",
+    "实际输入的目标文本(每句):": "实际输入的目标文本(每句):",
+    "实际输入的目标文本:": "实际输入的目标文本:",
+    "导出文件格式": "导出文件格式",
+    "已关闭": "已关闭",
+    "已完成": "已完成",
+    "已开启": "已开启",
+    "并行推理": "并行推理",
+    "并行推理模式已关闭": "并行推理模式已关闭",
+    "并行推理模式已开启": "并行推理模式已开启",
+    "开启": "开启",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "开启无参考文本模式。不填参考文本亦相当于开启。",
+    "微调训练": "微调训练",
+    "怎么切": "怎么切",
+    "总训练轮数total_epoch": "总训练轮数total_epoch",
+    "总训练轮数total_epoch，不建议太高": "总训练轮数total_epoch，不建议太高",
+    "指定输出主人声文件夹": "指定输出主人声文件夹",
+    "指定输出非主人声文件夹": "指定输出非主人声文件夹",
+    "按中文句号。切": "按中文句号。切",
+    "按标点符号切": "按标点符号切",
+    "按英文句号.切": "按英文句号.切",
+    "推理": "推理",
+    "推理设置": "推理设置",
+    "提取文本Bert特征": "提取文本Bert特征",
+    "数据分桶(并行推理时会降低一点计算量)": "数据分桶(并行推理时会降低一点计算量)",
+    "数据类型精度": "数据类型精度",
+    "文本分词与特征提取": "文本分词与特征提取",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。",
+    "文本模块学习率权重": "文本模块学习率权重",
+    "施工中，请静候佳音": "施工中，请静候佳音",
+    "日文": "日文",
+    "日英混合": "日英混合",
+    "是否仅保存最新的权重文件以节省硬盘空间": "是否仅保存最新的权重文件以节省硬盘空间",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "是否在每次保存时间点将最终小模型保存至weights文件夹",
+    "是否开启DPO训练选项(实验性)": "是否开启DPO训练选项(实验性)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "是否直接对上次合成结果调整语速和音色。防止随机性。",
+    "显卡信息": "显卡信息",
+    "未下载模型": "未下载模型",
+    "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.": "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.",
+    "标注文件路径 (含文件后缀 *.list)": "标注文件路径 (含文件后缀 *.list)",
+    "模型": "模型",
+    "模型分为三类：": "模型分为三类：",
+    "模型切换": "模型切换",
+    "每张显卡的batch_size": "每张显卡的batch_size",
+    "版本": "版本",
+    "粤英混合": "粤英混合",
+    "粤语": "粤语",
+    "终止合成": "终止合成",
+    "缺少Hubert数据集": "缺少Hubert数据集",
+    "缺少语义数据集": "缺少语义数据集",
+    "缺少音素数据集": "缺少音素数据集",
+    "缺少音频数据集": "缺少音频数据集",
+    "英文": "英文",
+    "训练集格式化一键三连": "训练集格式化一键三连",
+    "训练集格式化工具": "训练集格式化工具",
+    "语义Token提取": "语义Token提取",
+    "语速": "语速",
+    "语速调整，高为更快": "语速调整，高为更快",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "语速调节不支持分桶处理，已自动关闭分桶处理",
+    "语音切分": "语音切分",
+    "语音切分工具": "语音切分工具",
+    "语音文本校对标注工具": "语音文本校对标注工具",
+    "语音自监督特征提取": "语音自监督特征提取",
+    "语音识别": "语音识别",
+    "语音识别工具": "语音识别工具",
+    "语音降噪": "语音降噪",
+    "语音降噪工具": "语音降噪工具",
+    "请上传3~10秒内参考音频，超过会报错！": "请上传3~10秒内参考音频，超过会报错！",
+    "请上传参考音频": "请上传参考音频",
+    "请填入推理文本": "请填入推理文本",
+    "请填入正确的List路径": "请填入正确的List路径",
+    "请填入正确的音频文件夹路径": "请填入正确的音频文件夹路径",
+    "请输入有效文本": "请输入有效文本",
+    "路径不存在,使用默认配置": "路径不存在,使用默认配置",
+    "路径不能为空": "路径不能为空",
+    "路径错误": "路径错误",
+    "转换": "转换",
+    "辅参考音频(可选多个，或不选)": "辅参考音频(可选多个，或不选)",
+    "输入待处理音频文件夹路径": "输入待处理音频文件夹路径",
+    "输入文件夹路径": "输入文件夹路径",
+    "输入路径不存在": "输入路径不存在",
+    "输入路径存在但不可用": "输入路径存在但不可用",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "输出logs/实验名目录下应有23456开头的文件和文件夹",
+    "输出信息": "输出信息",
+    "输出文件夹路径": "输出文件夹路径",
+    "输出的语音": "输出的语音",
+    "运行中": "运行中",
+    "进度": "进度",
+    "进程已终止": "进程已终止",
+    "进程输出信息": "进程输出信息",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试",
+    "重复惩罚": "重复惩罚",
+    "随机种子": "随机种子",
+    "需先终止才能开启下一次任务": "需先终止才能开启下一次任务",
+    "需要合成的切分前文本": "需要合成的切分前文本",
+    "需要合成的文本": "需要合成的文本",
+    "需要合成的文本的语种": "需要合成的文本的语种",
+    "需要合成的语种": "需要合成的语种",
+    "韩文": "韩文",
+    "韩英混合": "韩英混合",
+    "音频加载失败": "音频加载失败",
+    "音频文件不存在，跳过：": "音频文件不存在，跳过：",
+    "音频标注WebUI": "音频标注WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "音频自动切分输入路径，可文件可文件夹",
+    "音频超分中": "音频超分中",
+    "预训练GPT模型路径": "预训练GPT模型路径",
+    "预训练SSL模型路径": "预训练SSL模型路径",
+    "预训练SoVITS-D模型路径": "预训练SoVITS-D模型路径",
+    "预训练SoVITS-G模型路径": "预训练SoVITS-G模型路径",
+    "预训练中文BERT模型路径": "预训练中文BERT模型路径"
+}
diff --git a/tools/i18n/locale/zh_HK.json b/tools/i18n/locale/zh_HK.json
new file mode 100644
index 0000000000000000000000000000000000000000..850b8328ee9197d92b78626c8680c0dd9be0ac26
--- /dev/null
+++ b/tools/i18n/locale/zh_HK.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb):對於雙通道混響是最佳選擇，但不能去除單通道混響；",
+    "(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。": "(234)DeEcho: 去除延遲效果。Aggressive 比 Normal 去除得更徹底，DeReverb 額外去除混響，可去除單聲道混響，但對高頻重的板式混響去不乾淨。",
+    "*实验/模型名": "*實驗/模型名",
+    "*文本标注文件": "*文本標注文件",
+    "*训练集音频文件目录": "*訓練集音頻文件目錄",
+    "*请上传并填写参考信息": "*請上傳並填寫參考信息",
+    "*请填写需要合成的目标文本和语种模式": "請填寫需要合成的目標文本和語言模式",
+    ".限制范围越小判别效果越好。": ".限制范围越小判别效果越好。",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1、DeEcho-DeReverb 模型的處理時間是另外兩個 DeEcho 模型的接近兩倍；",
+    "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；": "1、保留人聲：不帶和聲的音頻選這個，對主人聲保留比HP5更好。內置HP2和HP3兩個模型，HP3可能輕微漏伴奏但對主人聲保留比HP2稍微好一點點；",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-變聲",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverb 模型的處理時間挺慢的；",
+    "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；": "2、僅保留主人聲：帶和聲的音頻選這個，對主人聲可能有削弱。內置HP5一個模型；",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3、個人推薦的最乾淨的配置是先 MDX-Net 再 DeEcho-Aggressive。",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：",
+    "ASR 模型": "ASR 模型",
+    "ASR 模型尺寸": "ASR 模型尺寸",
+    "ASR 语言设置": "ASR 語言設置",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT 訓練: 模型權重檔案在 GPT_weights/",
+    "GPT模型列表": "GPT模型列表",
+    "GPT训练": "GPT訓練",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT 采样参数（无参考文本时不要太低。不懂就用默认）：",
+    "GPU卡号,只能填1个整数": "GPU卡號,只能填1個整數",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU卡號以-分割，每個卡號一個進程",
+    "LoRA秩": "LoRA秩",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 底模缺失，無法加載相應 LoRA 權重",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS 訓練: 模型權重檔案在 SoVITS_weights/",
+    "SoVITS模型列表": "SoVITS模型列表",
+    "SoVITS训练": "SoVITS訓練",
+    "TTS推理WebUI": "TTS推理WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5人聲伴奏分離&去混響去延遲工具",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:混多少比例歸一化後音頻進來",
+    "batch_size": "批次大小",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:怎麼算音量曲線，越小精度越大計算量越高（不是精度越大效果越好）",
+    "max:归一化后最大值多少": "max:歸一化後最大值多少",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:切完後靜音最多留多長",
+    "min_interval:最短切割间隔": "min_interval:最短切割間隔",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:每段最小多長，如果第一段太短一直和後面段連起來直到超過這個值",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:音量小於這個值視作靜音的備選切割點",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3暫不支持該模式，使用了會報錯。",
+    "v3输出如果觉得闷可以试试开超分": "v3輸出如果覺得悶可以試試開超分",
+    "不切": "不切",
+    "中文": "中文",
+    "中文教程文档": "中文教程文檔",
+    "中英混合": "中英混合",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "主參考音頻（請上傳3~10秒內參考音頻，超過會報錯！）",
+    "主参考音频的文本": "主參考音頻的文本",
+    "主参考音频的语种": "主參考音頻的語種",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "也可批量输入音频文件, 二选一, 优先读文件夹",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "人聲伴奏分離批量處理， 使用UVR5模型。",
+    "人声分离WebUI": "人聲分離WebUI",
+    "人声提取激进程度": "人聲提取激進程度",
+    "以下文件或文件夹不存在": "沒有此文件或文件夾",
+    "以下模型不存在:": "以下模型不存在:",
+    "伴奏人声分离&去混响&去回声": "伴奏人聲分離&去混響&去回聲",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "你沒有下載超分模型的參數，因此不進行超分。如想超分請先參照教程把文件下載好",
+    "使用无参考文本模式时建议使用微调的GPT": "使用無參考文本模式時建議使用微調的GPT",
+    "保存频率save_every_epoch": "保存頻率save_every_epoch",
+    "保持随机": "保持隨機",
+    "关闭": "關閉",
+    "凑50字一切": "湊50字一切",
+    "凑四句一切": "湊四句一切",
+    "分桶处理模式已关闭": "分桶處理模式已關閉",
+    "分桶处理模式已开启": "分桶處理模式已開啟",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "分段返回模式不支持分桶處理，已自動關閉分桶處理",
+    "分段返回模式已开启": "分段返回模式已開啟",
+    "分段间隔(秒)": "分段間隔（秒）",
+    "分段间隔过小，已自动设置为0.01": "分段間隔過小，已自動設置為0.01",
+    "切分": "切分",
+    "切分后文本": "切分後文本",
+    "切分后的子音频的输出根目录": "切分後的子音頻的輸出根目錄",
+    "切分文本": "切分文本",
+    "切割使用的进程数": "切割使用的進程數",
+    "刷新模型路径": "刷新模型路徑",
+    "前端处理后的文本(每句):": "前端處理後的文本(每句):",
+    "前置数据集获取工具": "前置數據集獲取工具",
+    "占用中": "佔用中",
+    "去混响/去延迟，附：": "去混響/去延遲，附",
+    "参考音频在3~10秒范围外，请更换！": "參考音頻在3~10秒範圍外，請更換！",
+    "参考音频的文本": "參考音頻的文本",
+    "参考音频的语种": "參考音頻的語種",
+    "句间停顿秒数": "句間停頓秒數",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "可選項：通過拖曳多個文件上傳多個參考音頻（建議同性），平均融合他們的音色。如不填寫此項，音色由左側單個參考音頻控制。如是微調模型，建議參考音頻全部在微調訓練集音色內，底模不用管。",
+    "合成语音": "合成語音",
+    "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)。": "合格的文件夾路徑格式舉例： E:\\codes\\py39\\vits_vc_gpu\\白鷺霜華測試樣例（去文件管理器地址欄拷就行了）。",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "後續將支持轉音素、手工修改音素、語音合成分步執行。",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "聽不清參考音頻說的啥(不曉得寫啥)可以開。開啟後無視填寫的參考文本。",
+    "启用并行推理版本": "啟用並行推理版本",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "填切割後音頻所在目錄！讀取的音頻文件完整路徑=該目錄-拼接-list文件裡波形對應的文件名（不是全路徑）。如果留空則使用.list文件裡的絕對全路徑。",
+    "多语种混合": "多語種混合",
+    "多语种混合(粤语)": "多語種混合 (粵語)",
+    "失败": "失敗",
+    "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.": "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文件. 詳見根目錄LICENSE.",
+    "实际输入的参考文本:": "實際輸入的參考文本:",
+    "实际输入的目标文本(切句后):": "實際輸入的目標文本（切句後）:",
+    "实际输入的目标文本(每句):": "實際輸入的目標文本（每句）:",
+    "实际输入的目标文本:": "實際輸入的目標文本:",
+    "导出文件格式": "導出檔格式",
+    "已关闭": "已關閉",
+    "已完成": "已完成",
+    "已开启": "已開啟",
+    "并行推理": "並行推理",
+    "并行推理模式已关闭": "並行推理模式已關閉",
+    "并行推理模式已开启": "並行推理模式已開啟",
+    "开启": "開啟",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "開啟無參考文本模式。不填參考文本亦相當於開啟。",
+    "微调训练": "微調訓練",
+    "怎么切": "怎麼切",
+    "总训练轮数total_epoch": "總訓練輪數total_epoch",
+    "总训练轮数total_epoch，不建议太高": "總訓練輪數total_epoch，不建議太高",
+    "指定输出主人声文件夹": "指定输出主人声文件夹",
+    "指定输出非主人声文件夹": "指定输出非主人声文件夹",
+    "按中文句号。切": "按中文句號。切",
+    "按标点符号切": "按標點符號切",
+    "按英文句号.切": "按英文句號.切",
+    "推理": "推理",
+    "推理设置": "推理設置",
+    "提取文本Bert特征": "提取文本BERT特徵",
+    "数据分桶(并行推理时会降低一点计算量)": "數據分桶（並行推理時會降低一點計算量）",
+    "数据类型精度": "數據類型精度",
+    "文本分词与特征提取": "文本分詞與特徵提取",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "文本切分工具。太長的文本合成出來效果不一定好，所以太長建議先切。合成會根據文本的換行分開合成再拼起來。",
+    "文本模块学习率权重": "文本模塊學習率權重",
+    "施工中，请静候佳音": "施工中，請靜候佳音",
+    "日文": "日文",
+    "日英混合": "日英混合",
+    "是否仅保存最新的权重文件以节省硬盘空间": "是否僅保存最新的權重文件以節省硬碟空間",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "是否在每次保存時間點將最終小模型保存至weights文件夾",
+    "是否开启DPO训练选项(实验性)": "是否開啟DPO訓練選項(實驗性)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "是否直接對上次合成結果調整語速和音色，以防止隨機性。",
+    "显卡信息": "顯卡信息",
+    "未下载模型": "未下載模型",
+    "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.": "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.",
+    "标注文件路径 (含文件后缀 *.list)": "標註文件路徑 (含文件後綴 *.list)",
+    "模型": "模型",
+    "模型分为三类：": "模型分為三類：",
+    "模型切换": "模型切換",
+    "每张显卡的batch_size": "每張顯卡的batch_size",
+    "版本": "版本",
+    "粤英混合": "粵英混合",
+    "粤语": "粵語",
+    "终止合成": "終止合成",
+    "缺少Hubert数据集": "缺少Hubert數據集",
+    "缺少语义数据集": "缺少語義數據集",
+    "缺少音素数据集": "缺少音素數據集",
+    "缺少音频数据集": "缺少音頻數據集",
+    "英文": "英文",
+    "训练集格式化一键三连": "訓練集格式化一鍵三連",
+    "训练集格式化工具": "訓練集格式化工具",
+    "语义Token提取": "語義Token提取",
+    "语速": "語速",
+    "语速调整，高为更快": "調整語速，高為更快",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "語速調節不支持分桶處理，已自動關閉分桶處理",
+    "语音切分": "語音切分",
+    "语音切分工具": "語音切分工具",
+    "语音文本校对标注工具": "語音文本校對標註工具",
+    "语音自监督特征提取": "語音自監督特徵提取",
+    "语音识别": "語音識別",
+    "语音识别工具": "語音識別工具",
+    "语音降噪": "語音降噪",
+    "语音降噪工具": "語音降噪工具",
+    "请上传3~10秒内参考音频，超过会报错！": "請上傳3~10秒內參考音頻，超過會報錯！",
+    "请上传参考音频": "請上傳參考音頻",
+    "请填入推理文本": "請填入推理文本",
+    "请填入正确的List路径": "請填寫正確的列表路徑",
+    "请填入正确的音频文件夹路径": "請填寫正確的音頻文件夾路徑",
+    "请输入有效文本": "請輸入有效文本",
+    "路径不存在,使用默认配置": "路徑不存在，使用默認配置",
+    "路径不能为空": "路徑不應該為空",
+    "路径错误": "路徑錯誤",
+    "转换": "轉換",
+    "辅参考音频(可选多个，或不选)": "輔參考音頻（可選多個，或不選）",
+    "输入待处理音频文件夹路径": "輸入待處理音頻資料夾路徑",
+    "输入文件夹路径": "輸入文件夾路徑",
+    "输入路径不存在": "輸入路徑不存在",
+    "输入路径存在但不可用": "輸入路徑存在但不可用",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "輸出logs/實驗名目錄下應有23456開頭的文件和文件夾",
+    "输出信息": "輸出訊息",
+    "输出文件夹路径": "輸出文件夾路徑",
+    "输出的语音": "輸出的語音",
+    "运行中": "運行中",
+    "进度": "進度",
+    "进程已终止": "進程已終止",
+    "进程输出信息": "進程輸出信息",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "選擇訓練完存放在SoVITS_weights和GPT_weights下的模型。默認的一個是底模，體驗5秒Zero Shot TTS用。",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "採樣步數，如果覺得電，提高試試，如果覺得慢，降低試試",
+    "重复惩罚": "重複懲罰",
+    "随机种子": "隨機種子",
+    "需先终止才能开启下一次任务": "需先終止才能開啟下一次任務",
+    "需要合成的切分前文本": "需要合成的切分前文本",
+    "需要合成的文本": "需要合成的文本",
+    "需要合成的文本的语种": "需要合成的文本的語種",
+    "需要合成的语种": "需要合成的語種",
+    "韩文": "韓文",
+    "韩英混合": "韓英混合",
+    "音频加载失败": "無法加載音頻",
+    "音频文件不存在，跳过：": "音頻檔案不存在，跳過：",
+    "音频标注WebUI": "音頻標註WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "音頻自動切分輸入路徑，可文件可文件夾",
+    "音频超分中": "音頻超分中",
+    "预训练GPT模型路径": "預訓練GPT模型路徑",
+    "预训练SSL模型路径": "預訓練SSL模型路徑",
+    "预训练SoVITS-D模型路径": "預訓練SoVITS-D模型路徑",
+    "预训练SoVITS-G模型路径": "預訓練SoVITS-G模型路徑",
+    "预训练中文BERT模型路径": "預訓練中文BERT模型路徑"
+}
diff --git a/tools/i18n/locale/zh_SG.json b/tools/i18n/locale/zh_SG.json
new file mode 100644
index 0000000000000000000000000000000000000000..9ea41ecd29f4ea00ee3cae69aa19de4894f5f9cf
--- /dev/null
+++ b/tools/i18n/locale/zh_SG.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb)：對於雙通道混響是最好的選擇，不能去除單通道混響；",
+    "(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。": "(234)DeEcho：去除延遲效果。Aggressive 比 Normal 去除得更徹底，DeReverb 額外去除混響，可去除單聲道混響，但是對高頻重的板式混響去不乾淨。",
+    "*实验/模型名": "*實驗/模型名",
+    "*文本标注文件": "*文本標註文件",
+    "*训练集音频文件目录": "*訓練集音頻文件目錄",
+    "*请上传并填写参考信息": "*請上傳並填寫參考信息",
+    "*请填写需要合成的目标文本和语种模式": "請填寫需要合成的目標文本和語言模式",
+    ".限制范围越小判别效果越好。": ".限制范围越小判别效果越好。",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1、DeEcho-DeReverb 模型的耗時是另外兩個 DeEcho 模型的接近兩倍；",
+    "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；": "1、保留人聲：不帶和聲的音頻選這個，對主人聲保留比HP5更好。內置HP2和HP3兩個模型，HP3可能輕微漏伴奏但對主人聲保留比HP2稍微好一丁點；",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-變聲",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverb 模型的處理時間挺慢的；",
+    "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；": "2、僅保留主人聲：帶和聲的音頻選這個，對主人聲可能有削弱。內置HP5一個模型；",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3、個人推薦的最乾淨的配置是先 MDX-Net 再 DeEcho-Aggressive。",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：",
+    "ASR 模型": "ASR 模型",
+    "ASR 模型尺寸": "ASR 模型尺寸",
+    "ASR 语言设置": "ASR 語言設定",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT 训练: 模型权重文件在 GPT_weights/ 目錄下",
+    "GPT模型列表": "GPT模型列表",
+    "GPT训练": "GPT訓練",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT 采样参数（无参考文本时不要太低。不懂就用默认）：",
+    "GPU卡号,只能填1个整数": "GPU卡號,只能填1個整數",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU卡號以-分割，每個卡號一個進程",
+    "LoRA秩": "LoRA秩",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 底模缺失，無法加載相應 LoRA 權重",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS 訓練: 模型權重文件在 SoVITS_weights/ 目錄下",
+    "SoVITS模型列表": "SoVITS模型列表",
+    "SoVITS训练": "SoVITS訓練",
+    "TTS推理WebUI": "TTS推理WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5人聲伴奏分離&去混響去延遲工具",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:混多少比例歸一化後音頻進來",
+    "batch_size": "批次大小",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:怎麼算音量曲線，越小精度越大計算量越高（不是精度越大效果越好）",
+    "max:归一化后最大值多少": "max:歸一化後最大值多少",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:切完後靜音最多留多長",
+    "min_interval:最短切割间隔": "min_interval:最短切割間隔",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:每段最小多長，如果第一段太短一直和後面段連起來直到超過這個值",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:音量小於這個值視作靜音的備選切割點",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3暫不支援該模式，使用了會報錯。",
+    "v3输出如果觉得闷可以试试开超分": "v3輸出如果覺得悶可以試試開超分",
+    "不切": "不切",
+    "中文": "中文",
+    "中文教程文档": "中文教程文檔",
+    "中英混合": "中英混合",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "主參考音檔（請上傳3~10秒內參考音檔，超過會報錯！）",
+    "主参考音频的文本": "主參考音檔的文本",
+    "主参考音频的语种": "主參考音檔的語種",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "也可批量输入音频文件, 二选一, 优先读文件夹",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "人聲伴奏分離批量處理， 使用UVR5模型。",
+    "人声分离WebUI": "人聲分離WebUI",
+    "人声提取激进程度": "人聲提取激進程度",
+    "以下文件或文件夹不存在": "沒有此文件或文件夾",
+    "以下模型不存在:": "以下模型不存在",
+    "伴奏人声分离&去混响&去回声": "伴奏人聲分離&去混響&去回聲",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "你沒有下載超分模型的參數，因此不進行超分。如想超分請先參照教程把文件下載好",
+    "使用无参考文本模式时建议使用微调的GPT": "使用無參考文本模式時建議使用微調的GPT",
+    "保存频率save_every_epoch": "保存頻率save_every_epoch",
+    "保持随机": "保持隨機",
+    "关闭": "關閉",
+    "凑50字一切": "湊50字一切",
+    "凑四句一切": "湊四句一切",
+    "分桶处理模式已关闭": "分桶處理模式已關閉",
+    "分桶处理模式已开启": "分桶處理模式已開啟",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "分段返回模式不支援分桶處理，已自動關閉分桶處理",
+    "分段返回模式已开启": "分段返回模式已開啟",
+    "分段间隔(秒)": "分段間隔（秒）",
+    "分段间隔过小，已自动设置为0.01": "分段間隔過小，已自動設置為0.01",
+    "切分": "切分",
+    "切分后文本": "切分後文本",
+    "切分后的子音频的输出根目录": "切分後的子音頻的輸出根目錄",
+    "切分文本": "切分文本",
+    "切割使用的进程数": "切割使用的進程數",
+    "刷新模型路径": "刷新模型路徑",
+    "前端处理后的文本(每句):": "前端處理後的文本(每句):",
+    "前置数据集获取工具": "前置數據集獲取工具",
+    "占用中": "占用中",
+    "去混响/去延迟，附：": "去混響/去延遲，附：",
+    "参考音频在3~10秒范围外，请更换！": "參考音頻在3~10秒範圍外，請更換！",
+    "参考音频的文本": "參考音頻的文本",
+    "参考音频的语种": "參考音頻的語種",
+    "句间停顿秒数": "句間停頓秒數",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "可選項：通過拖曳多個文件上傳多個參考音頻（建議同性），平均融合他們的音色。如不填寫此項，音色由左側單個參考音頻控制。如是微調模型，建議參考音頻全部在微調訓練集音色內，底模不用管。",
+    "合成语音": "合成語音",
+    "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)。": "合格的資料夾路徑格式舉例： E:\\codes\\py39\\vits_vc_gpu\\白鷺霜華測試範例(去文件管理器地址欄拷就行了)。",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "後續將支援轉音素、手工修改音素、語音合成分步執行。",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "聽不清參考音頻說的啥(不曉得寫啥)可以開。開啟後無視填寫的參考文本。",
+    "启用并行推理版本": "啟用並行推理版本",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "填切割後音頻所在目錄！讀取的音頻檔案完整路徑=該目錄-拼接-list檔案裡波形對應的檔案名（不是全路徑）。如果留空則使用.list檔案裡的絕對全路徑。",
+    "多语种混合": "多語種混合",
+    "多语种混合(粤语)": "多語種混合 (粵語)",
+    "失败": "失敗",
+    "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.": "如不認可該條款, 則不能使用或引用軟體包內任何代碼和文件. 詳見根目錄LICENSE.",
+    "实际输入的参考文本:": "實際輸入的參考文本:",
+    "实际输入的目标文本(切句后):": "實際輸入的目標文本(切句後):",
+    "实际输入的目标文本(每句):": "實際輸入的目標文本(每句):",
+    "实际输入的目标文本:": "實際輸入的目標文本:",
+    "导出文件格式": "導出檔格式",
+    "已关闭": "已關閉",
+    "已完成": "已完成",
+    "已开启": "已開啟",
+    "并行推理": "並行推理",
+    "并行推理模式已关闭": "並行推理模式已關閉",
+    "并行推理模式已开启": "並行推理模式已開啟",
+    "开启": "開啟",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "開啟無參考文本模式。不填參考文本亦相當於開啟。",
+    "微调训练": "微調訓練",
+    "怎么切": "怎麼切",
+    "总训练轮数total_epoch": "總訓練輪數total_epoch",
+    "总训练轮数total_epoch，不建议太高": "總訓練輪數total_epoch，不建議太高",
+    "指定输出主人声文件夹": "指定输出主人声文件夹",
+    "指定输出非主人声文件夹": "指定输出非主人声文件夹",
+    "按中文句号。切": "按中文句號。切",
+    "按标点符号切": "按標點符號切",
+    "按英文句号.切": "按英文句號.切",
+    "推理": "推理",
+    "推理设置": "推理設定",
+    "提取文本Bert特征": "提取文本BERT特徵",
+    "数据分桶(并行推理时会降低一点计算量)": "資料分桶（並行推理時會降低一點計算量）",
+    "数据类型精度": "數據類型精度",
+    "文本分词与特征提取": "文本分詞與特徵提取",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "文本切分工具。太長的文本合成出來效果不一定好，所以太長建議先切。合成會根據文本的換行分開合成再拼起來。",
+    "文本模块学习率权重": "文本模塊學習率權重",
+    "施工中，请静候佳音": "施工中，請靜候佳音",
+    "日文": "日文",
+    "日英混合": "日英混合",
+    "是否仅保存最新的权重文件以节省硬盘空间": "是否僅保存最新的權重文件以節省硬碟空間",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "是否在每次保存時間點將最終小模型保存至weights文件夾",
+    "是否开启DPO训练选项(实验性)": "是否開啟DPO訓練選項(實驗性)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "是否直接對上次合成結果調整語速和音色，以防止隨機性。",
+    "显卡信息": "顯卡資訊",
+    "未下载模型": "未下載模型",
+    "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.": "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.",
+    "标注文件路径 (含文件后缀 *.list)": "標註文件路徑 (含文件後綴 *.list)",
+    "模型": "模型",
+    "模型分为三类：": "模型分為三類：",
+    "模型切换": "模型切換",
+    "每张显卡的batch_size": "每張顯卡的batch_size",
+    "版本": "版本",
+    "粤英混合": "粵英混合",
+    "粤语": "粵語",
+    "终止合成": "終止合成",
+    "缺少Hubert数据集": "缺少Hubert數據集",
+    "缺少语义数据集": "缺少語義數據集",
+    "缺少音素数据集": "缺少音素數據集",
+    "缺少音频数据集": "缺少音頻數據集",
+    "英文": "英文",
+    "训练集格式化一键三连": "訓練集格式化一鍵三連",
+    "训练集格式化工具": "訓練集格式化工具",
+    "语义Token提取": "語義Token提取",
+    "语速": "語速",
+    "语速调整，高为更快": "調整語速，高為更快",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "語速調節不支援分桶處理，已自動關閉分桶處理",
+    "语音切分": "語音切分",
+    "语音切分工具": "語音切分工具",
+    "语音文本校对标注工具": "語音文本校對標註工具",
+    "语音自监督特征提取": "語音自監督特徵提取",
+    "语音识别": "語音識別",
+    "语音识别工具": "語音識別工具",
+    "语音降噪": "語音降噪",
+    "语音降噪工具": "語音降噪工具",
+    "请上传3~10秒内参考音频，超过会报错！": "請上傳3~10秒內參考音頻，超過會報錯！",
+    "请上传参考音频": "請上傳參考音頻",
+    "请填入推理文本": "請填入推理文本",
+    "请填入正确的List路径": "請填寫正確的列表路徑",
+    "请填入正确的音频文件夹路径": "請填寫正確的音頻文件夾路徑",
+    "请输入有效文本": "請輸入有效文本",
+    "路径不存在,使用默认配置": "路徑不存在，使用預設配置",
+    "路径不能为空": "路徑不應該為空",
+    "路径错误": "路徑錯誤",
+    "转换": "轉換",
+    "辅参考音频(可选多个，或不选)": "輔參考音檔（可選多個，或不選）",
+    "输入待处理音频文件夹路径": "輸入待處理音頻資料夾路徑",
+    "输入文件夹路径": "輸入文件夾路徑",
+    "输入路径不存在": "輸入路徑不存在",
+    "输入路径存在但不可用": "輸入路徑存在但不可用",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "輸出logs/實驗名目錄下應有23456開頭的文件和文件夾",
+    "输出信息": "輸出訊息",
+    "输出文件夹路径": "輸出文件夾路徑",
+    "输出的语音": "輸出的語音",
+    "运行中": "運行中",
+    "进度": "進度",
+    "进程已终止": "進程已終止",
+    "进程输出信息": "進程輸出信息",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "選擇訓練完存放在SoVITS_weights和GPT_weights下的模型。默認的一個是底模，體驗5秒Zero Shot TTS用。",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "採樣步數，如果覺得電，提高試試，如果覺得慢，降低試試",
+    "重复惩罚": "重複懲罰",
+    "随机种子": "隨機種子",
+    "需先终止才能开启下一次任务": "需先終止才能開啟下一次任務",
+    "需要合成的切分前文本": "需要合成的切分前文本",
+    "需要合成的文本": "需要合成的文本",
+    "需要合成的文本的语种": "需要合成的文本的語種",
+    "需要合成的语种": "需要合成的語種",
+    "韩文": "韓文",
+    "韩英混合": "韓英混合",
+    "音频加载失败": "無法加載音頻",
+    "音频文件不存在，跳过：": "音檔不存在，跳過：",
+    "音频标注WebUI": "音頻標註WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "音頻自動切分輸入路徑，可文件可文件夾",
+    "音频超分中": "音頻超分中",
+    "预训练GPT模型路径": "預訓練GPT模型路徑",
+    "预训练SSL模型路径": "預訓練SSL模型路徑",
+    "预训练SoVITS-D模型路径": "預訓練SoVITS-D模型路徑",
+    "预训练SoVITS-G模型路径": "預訓練SoVITS-G模型路徑",
+    "预训练中文BERT模型路径": "預訓練中文BERT模型路徑"
+}
diff --git a/tools/i18n/locale/zh_TW.json b/tools/i18n/locale/zh_TW.json
new file mode 100644
index 0000000000000000000000000000000000000000..70a3f2a95129f45c7b98496403d738ab80b33bcc
--- /dev/null
+++ b/tools/i18n/locale/zh_TW.json
@@ -0,0 +1,211 @@
+{
+    "(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；": "(1)MDX-Net(onnx_dereverb)：對於雙通道混響是最好的選擇，不能去除單通道混響；",
+    "(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。": "(234)DeEcho：去除延遲效果。Aggressive 比 Normal 去除得更徹底，DeReverb 額外去除混響，可去除單聲道混響，但是對高頻重的板式混響去不乾淨。",
+    "*实验/模型名": "*實驗/模型名",
+    "*文本标注文件": "*文本標注文件",
+    "*训练集音频文件目录": "*訓練集音頻文件目錄",
+    "*请上传并填写参考信息": "*請上傳並填寫參考資訊",
+    "*请填写需要合成的目标文本和语种模式": "請填寫需要合成的目標文本和語言模式",
+    ".限制范围越小判别效果越好。": ".限制范围越小判别效果越好。",
+    "1-GPT-SoVITS-TTS": "1-GPT-SoVITS-TTS",
+    "1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；": "1、DeEcho-DeReverb 模型的耗時是另外兩個 DeEcho 模型的接近兩倍；",
+    "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；": "1、保留人聲：不帶和聲的音頻選這個，對主人聲保留比HP5更好。內置HP2和HP3兩個模型，HP3可能輕微漏伴奏但對主人聲保留比HP2稍微好一丁點；",
+    "2-GPT-SoVITS-变声": "2-GPT-SoVITS-變聲",
+    "2、MDX-Net-Dereverb模型挺慢的；": "2、MDX-Net-Dereverb模型挺慢的；",
+    "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；": "2、僅保留主人聲：帶和聲的音頻選這個，對主人聲可能有削弱。內置HP5一個模型；",
+    "3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。": "3、個人推薦的最乾淨的配置是先 MDX-Net 再 DeEcho-Aggressive。",
+    "3、去混响、去延迟模型（by FoxJoy）：": "3、去混響、去延遲模型（by FoxJoy）：",
+    "ASR 模型": "ASR 模型",
+    "ASR 模型尺寸": "ASR 模型尺寸",
+    "ASR 语言设置": "ASR 語言設置",
+    "GPT 训练: 模型权重文件在 GPT_weights/": "GPT 訓練: 模型權重文件在 GPT_weights/",
+    "GPT模型列表": "GPT模型列表",
+    "GPT训练": "GPT訓練",
+    "GPT采样参数(无参考文本时不要太低。不懂就用默认)：": "GPT 采样参数（无参考文本时不要太低。不懂就用默认）：",
+    "GPU卡号,只能填1个整数": "GPU卡號,只能填1個整數",
+    "GPU卡号以-分割，每个卡号一个进程": "GPU卡號以-分割，每個卡號一個進程",
+    "LoRA秩": "LoRA階",
+    "SoVITS V3 底模缺失，无法加载相应 LoRA 权重": "SoVITS V3 底模缺失，無法加載相應 LoRA 權重",
+    "SoVITS 训练: 模型权重文件在 SoVITS_weights/": "SoVITS 訓練: 模型權重文件在 SoVITS_weights/",
+    "SoVITS模型列表": "SoVITS模型列表",
+    "SoVITS训练": "SoVITS訓練",
+    "TTS推理WebUI": "TTS推理WebUI",
+    "UVR5人声伴奏分离&去混响去延迟工具": "UVR5人聲伴奏分離&去混響去延遲工具",
+    "alpha_mix:混多少比例归一化后音频进来": "alpha_mix:混多少比例歸一化後音頻進來",
+    "batch_size": "批次大小",
+    "hop_size:怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）": "hop_size:怎麼算音量曲線，越小精度越大計算量越高（不是精度越大效果越好）",
+    "max:归一化后最大值多少": "max:歸一化後最大值多少",
+    "max_sil_kept:切完后静音最多留多长": "max_sil_kept:切完後靜音最多留多長",
+    "min_interval:最短切割间隔": "min_interval:最短切割間隔",
+    "min_length:每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值": "min_length:每段最小多長，如果第一段太短一直和後面段連起來直到超過這個值",
+    "temperature": "temperature",
+    "threshold:音量小于这个值视作静音的备选切割点": "threshold:音量小於這個值視作靜音的備選切割點",
+    "top_k": "top_k",
+    "top_p": "top_p",
+    "v3暂不支持该模式，使用了会报错。": "v3暫不支援該模式，使用了會報錯。",
+    "v3输出如果觉得闷可以试试开超分": "v3輸出如果覺得悶可以試試開超分",
+    "不切": "不切",
+    "中文": "中文",
+    "中文教程文档": "中文教程文檔",
+    "中英混合": "中英混合",
+    "主参考音频(请上传3~10秒内参考音频，超过会报错！)": "主參考音檔（請上傳3~10秒內參考音檔，超過會報錯！）",
+    "主参考音频的文本": "主參考音檔的文字",
+    "主参考音频的语种": "主參考音檔的語種",
+    "也可批量输入音频文件, 二选一, 优先读文件夹": "也可批量输入音频文件, 二选一, 优先读文件夹",
+    "人声伴奏分离批量处理， 使用UVR5模型。": "人聲伴奏分離批量處理， 使用UVR5模型。",
+    "人声分离WebUI": "人聲分離WebUI",
+    "人声提取激进程度": "人聲提取激進程度",
+    "以下文件或文件夹不存在": "沒有此文件或文件夾",
+    "以下模型不存在:": "#以下模型不存在",
+    "伴奏人声分离&去混响&去回声": "伴奏人聲分離&去混響&去回聲",
+    "你没有下载超分模型的参数，因此不进行超分。如想超分请先参照教程把文件下载好": "你沒有下載超分模型的參數，因此不進行超分。如想超分請先參照教程把文件下載好",
+    "使用无参考文本模式时建议使用微调的GPT": "使用無參考文本模式時建議使用微調的GPT",
+    "保存频率save_every_epoch": "保存頻率save_every_epoch",
+    "保持随机": "保持隨機",
+    "关闭": "關閉",
+    "凑50字一切": "湊50字一切",
+    "凑四句一切": "湊四句一切",
+    "分桶处理模式已关闭": "分桶處理模式已關閉",
+    "分桶处理模式已开启": "分桶處理模式已開啟",
+    "分段返回模式不支持分桶处理，已自动关闭分桶处理": "分段返回模式不支援分桶處理，已自動關閉分桶處理",
+    "分段返回模式已开启": "分段返回模式已開啟",
+    "分段间隔(秒)": "分段間隔（秒）",
+    "分段间隔过小，已自动设置为0.01": "分段間隔過小，已自動設定為0.01",
+    "切分": "切分",
+    "切分后文本": "切分後文字",
+    "切分后的子音频的输出根目录": "切分後的子音頻的輸出根目錄",
+    "切分文本": "切分文字",
+    "切割使用的进程数": "切割使用的進程數",
+    "刷新模型路径": "刷新模型路徑",
+    "前端处理后的文本(每句):": "前端處理後的文本(每句):",
+    "前置数据集获取工具": "前置數據集獲取工具",
+    "占用中": "佔用中",
+    "去混响/去延迟，附：": "去混響/去延遲，附：",
+    "参考音频在3~10秒范围外，请更换！": "參考音頻在3~10秒範圍外，請更換！",
+    "参考音频的文本": "參考音頻的文本",
+    "参考音频的语种": "參考音頻的語種",
+    "句间停顿秒数": "句間停頓秒數",
+    "可选项：通过拖拽多个文件上传多个参考音频（建议同性），平均融合他们的音色。如不填写此项，音色由左侧单个参考音频控制。如是微调模型，建议参考音频全部在微调训练集音色内，底模不用管。": "可選項：通過拖曳多個文件上傳多個參考音頻（建議同性），平均融合他們的音色。如不填寫此項，音色由左側單個參考音頻控制。如是微調模型，建議參考音頻全部在微調訓練集音色內，底模不用管。",
+    "合成语音": "合成語音",
+    "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)。": "合格的資料夾路徑格式舉例： E:\\codes\\py39\\vits_vc_gpu\\白鷺霜華測試範例(去文件管理器地址欄拷就行了)。",
+    "后续将支持转音素、手工修改音素、语音合成分步执行。": "後續將支援轉音素、手工修改音素、語音合成分步執行。",
+    "听不清参考音频说的啥(不晓得写啥)可以开。开启后无视填写的参考文本。": "聽不清參考音頻說的啥(不曉得寫啥)可以開，開啟後無視填寫的參考文本。",
+    "启用并行推理版本": "啟用並行推理版本",
+    "填切割后音频所在目录！读取的音频文件完整路径=该目录-拼接-list文件里波形对应的文件名（不是全路径）。如果留空则使用.list文件里的绝对全路径。": "填切割後音頻所在目錄！讀取的音頻檔案完整路徑=該目錄-拼接-list檔案裡波形對應的檔案名（不是全路徑）。如果留空則使用.list檔案裡的絕對全路徑。",
+    "多语种混合": "多語種混合",
+    "多语种混合(粤语)": "多語種混合 (粵語)",
+    "失败": "失敗",
+    "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.": "如不認可該條款，則不能使用或引用軟體包內任何代碼和文件。詳見根目錄LICENSE",
+    "实际输入的参考文本:": "實際輸入的參考文本:",
+    "实际输入的目标文本(切句后):": "實際輸入的目標文本(切句後):",
+    "实际输入的目标文本(每句):": "實際輸入的目標文本(每句):",
+    "实际输入的目标文本:": "實際輸入的目標文本:",
+    "导出文件格式": "導出檔格式",
+    "已关闭": "已關閉",
+    "已完成": "已完成",
+    "已开启": "已開啟",
+    "并行推理": "並行推理",
+    "并行推理模式已关闭": "並行推理模式已關閉",
+    "并行推理模式已开启": "並行推理模式已開啟",
+    "开启": "開啟",
+    "开启无参考文本模式。不填参考文本亦相当于开启。": "開啟無參考文本模式。不填參考文本亦相當於開啟。",
+    "微调训练": "微調訓練",
+    "怎么切": "怎麼切",
+    "总训练轮数total_epoch": "總訓練輪數total_epoch",
+    "总训练轮数total_epoch，不建议太高": "總訓練輪數total_epoch，不建議太高",
+    "指定输出主人声文件夹": "指定输出主人声文件夹",
+    "指定输出非主人声文件夹": "指定输出非主人声文件夹",
+    "按中文句号。切": "按中文句號。切",
+    "按标点符号切": "按標點符號切",
+    "按英文句号.切": "按英文句號.切",
+    "推理": "推理",
+    "推理设置": "推理設定",
+    "提取文本Bert特征": "提取文字BERT特徵",
+    "数据分桶(并行推理时会降低一点计算量)": "資料分桶（並行推理時會降低一點計算量）",
+    "数据类型精度": "數據類型精度",
+    "文本分词与特征提取": "文本分詞與特徵提取",
+    "文本切分工具。太长的文本合成出来效果不一定好，所以太长建议先切。合成会根据文本的换行分开合成再拼起来。": "文字切分工具。太長的文字合成出來效果不一定好，所以太長建議先切。合成會根據文字的換行分開合成再拼起來。",
+    "文本模块学习率权重": "文本模塊學習率權重",
+    "施工中，请静候佳音": "施工中，請靜候佳音",
+    "日文": "日文",
+    "日英混合": "日英混合",
+    "是否仅保存最新的权重文件以节省硬盘空间": "是否僅保存最新的權重文件以節省硬盤空間",
+    "是否在每次保存时间点将最终小模型保存至weights文件夹": "是否在每次保存時間點將最終小模型保存至weights文件夾",
+    "是否开启DPO训练选项(实验性)": "是否開啟DPO訓練選項(實驗性)",
+    "是否直接对上次合成结果调整语速和音色。防止随机性。": "是否直接對上次合成結果調整語速和音色，以防止隨機性。",
+    "显卡信息": "顯卡資訊",
+    "未下载模型": "未下載模型",
+    "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.": "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。",
+    "标注文件路径 (含文件后缀 *.list)": "標註文件路徑 (含文件後綴 *.list)",
+    "模型": "模型",
+    "模型分为三类：": "模型分為三類：",
+    "模型切换": "模型切換",
+    "每张显卡的batch_size": "每張顯卡的batch_size",
+    "版本": "版本",
+    "粤英混合": "粵英混合",
+    "粤语": "粵語",
+    "终止合成": "終止合成",
+    "缺少Hubert数据集": "缺少Hubert數據集",
+    "缺少语义数据集": "缺少語義數據集",
+    "缺少音素数据集": "缺少音素數據集",
+    "缺少音频数据集": "缺少音頻數據集",
+    "英文": "英文",
+    "训练集格式化一键三连": "訓練集格式化一鍵三連",
+    "训练集格式化工具": "訓練集格式化工具",
+    "语义Token提取": "語義Token提取",
+    "语速": "語速",
+    "语速调整，高为更快": "調整語速，高為更快",
+    "语速调节不支持分桶处理，已自动关闭分桶处理": "語速調節不支援分桶處理，已自動關閉分桶處理",
+    "语音切分": "語音切分",
+    "语音切分工具": "語音切分工具",
+    "语音文本校对标注工具": "語音文本校對標註工具",
+    "语音自监督特征提取": "語音自監督特徵提取",
+    "语音识别": "語音識別",
+    "语音识别工具": "語音識別工具",
+    "语音降噪": "語音降噪",
+    "语音降噪工具": "語音降噪工具",
+    "请上传3~10秒内参考音频，超过会报错！": "請上傳3~10秒內參考音頻，超過會報錯！",
+    "请上传参考音频": "請上傳參考音頻",
+    "请填入推理文本": "請填入推理文本",
+    "请填入正确的List路径": "請填寫正確的列表路徑",
+    "请填入正确的音频文件夹路径": "請填寫正確的音頻文件夾路徑",
+    "请输入有效文本": "請輸入有效文本",
+    "路径不存在,使用默认配置": "路徑不存在，使用預設配置",
+    "路径不能为空": "路徑不應該為空",
+    "路径错误": "路徑錯誤",
+    "转换": "轉換",
+    "辅参考音频(可选多个，或不选)": "輔參考音檔（可選多個，或不選）",
+    "输入待处理音频文件夹路径": "輸入待處理音頻資料夾路徑",
+    "输入文件夹路径": "輸入文件夾路徑",
+    "输入路径不存在": "輸入路徑不存在",
+    "输入路径存在但不可用": "輸入路徑存在但不可用",
+    "输出logs/实验名目录下应有23456开头的文件和文件夹": "輸出logs/實驗名目錄下應有23456開頭的文件和文件夾",
+    "输出信息": "輸出訊息",
+    "输出文件夹路径": "輸出文件夾路徑",
+    "输出的语音": "輸出的語音",
+    "运行中": "運行中",
+    "进度": "進度",
+    "进程已终止": "進程已終止",
+    "进程输出信息": "進程輸出資訊",
+    "选择训练完存放在SoVITS_weights和GPT_weights下的模型。默认的一个是底模，体验5秒Zero Shot TTS用。": "選擇訓練完存放在SoVITS_weights和GPT_weights下的模型。默認的一個是底模，體驗5秒Zero Shot TTS用。",
+    "采样步数,如果觉得电,提高试试,如果觉得慢,降低试试": "採樣步數，如果覺得電，提高試試，如果覺得慢，降低試試",
+    "重复惩罚": "重複懲罰",
+    "随机种子": "隨機種子",
+    "需先终止才能开启下一次任务": "需先終止才能開啟下一次任務",
+    "需要合成的切分前文本": "需要合成的切分前文字",
+    "需要合成的文本": "需要合成的文本",
+    "需要合成的文本的语种": "需要合成的文字語種",
+    "需要合成的语种": "需要合成的語種",
+    "韩文": "韓文",
+    "韩英混合": "韓英混合",
+    "音频加载失败": "無法加載音頻",
+    "音频文件不存在，跳过：": "音檔不存在，跳過：",
+    "音频标注WebUI": "音頻標註WebUI",
+    "音频自动切分输入路径，可文件可文件夹": "音頻自動切分輸入路徑，可文件可文件夾",
+    "音频超分中": "音頻超分中",
+    "预训练GPT模型路径": "預訓練GPT模型路徑",
+    "预训练SSL模型路径": "預訓練SSL模型路徑",
+    "预训练SoVITS-D模型路径": "預訓練SoVITS-D模型路徑",
+    "预训练SoVITS-G模型路径": "預訓練SoVITS-G模型路徑",
+    "预训练中文BERT模型路径": "預訓練中文BERT模型路徑"
+}
diff --git a/tools/i18n/scan_i18n.py b/tools/i18n/scan_i18n.py
new file mode 100644
index 0000000000000000000000000000000000000000..0f12091deacedb430afd881472a6844a5592fa30
--- /dev/null
+++ b/tools/i18n/scan_i18n.py
@@ -0,0 +1,137 @@
+import ast
+import glob
+import json
+import os
+from collections import OrderedDict
+
+I18N_JSON_DIR: os.PathLike = os.path.join(os.path.dirname(os.path.relpath(__file__)), "locale")
+DEFAULT_LANGUAGE: str = "zh_CN"  # 默认语言
+TITLE_LEN: int = 60  # 标题显示长度
+KEY_LEN: int = 30  # 键名显示长度
+SHOW_KEYS: bool = False  # 是否显示键信息
+SORT_KEYS: bool = False  # 是否按全局键名写入文件
+
+
+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
+
+
+def scan_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 = []
+    print(" Scanning Files and Extracting i18n Strings ".center(TITLE_LEN, "="))
+    for filename in glob.iglob("**/*.py", recursive=True):
+        try:
+            with open(filename, "r", encoding="utf-8") as f:
+                code = f.read()
+                if "I18nAuto" in code:
+                    tree = ast.parse(code)
+                    i18n_strings = extract_i18n_strings(tree)
+                    print(f"{filename.ljust(KEY_LEN * 3 // 2)}: {len(i18n_strings)}")
+                    if SHOW_KEYS:
+                        print("\n".join([s for s in i18n_strings]))
+                    strings.extend(i18n_strings)
+        except Exception as e:
+            print(f"\033[31m[Failed] Error occur at {filename}: {e}\033[0m")
+
+    code_keys = set(strings)
+    print(f"{'Total Unique'.ljust(KEY_LEN * 3 // 2)}: {len(code_keys)}")
+    return code_keys
+
+
+def update_i18n_json(json_file, standard_keys):
+    standard_keys = sorted(standard_keys)
+    print(f" Process {json_file} ".center(TITLE_LEN, "="))
+    # 读取 JSON 文件
+    with open(json_file, "r", encoding="utf-8") as f:
+        json_data = json.load(f, object_pairs_hook=OrderedDict)
+    # 打印处理前的 JSON 条目数
+    len_before = len(json_data)
+    print(f"{'Total Keys'.ljust(KEY_LEN)}: {len_before}")
+    # 识别缺失的键并补全
+    miss_keys = set(standard_keys) - set(json_data.keys())
+    if len(miss_keys) > 0:
+        print(f"{'Missing Keys (+)'.ljust(KEY_LEN)}: {len(miss_keys)}")
+        for key in miss_keys:
+            if DEFAULT_LANGUAGE in json_file:
+                # 默认语言的键值相同.
+                json_data[key] = key
+            else:
+                # 其他语言的值设置为 #! + 键名以标注未被翻译.
+                json_data[key] = "#!" + key
+            if SHOW_KEYS:
+                print(f"{'Added Missing Key'.ljust(KEY_LEN)}: {key}")
+    # 识别多余的键并删除
+    diff_keys = set(json_data.keys()) - set(standard_keys)
+    if len(diff_keys) > 0:
+        print(f"{'Unused Keys  (-)'.ljust(KEY_LEN)}: {len(diff_keys)}")
+        for key in diff_keys:
+            del json_data[key]
+            if SHOW_KEYS:
+                print(f"{'Removed Unused Key'.ljust(KEY_LEN)}: {key}")
+    # 按键顺序排序
+    json_data = OrderedDict(
+        sorted(
+            json_data.items(),
+            key=lambda x: (
+                list(standard_keys).index(x[0])
+                if x[0] in standard_keys and not x[1].startswith("#!")
+                else len(json_data),
+            ),
+        )
+    )
+    # 打印处理后的 JSON 条目数
+    if len(miss_keys) != 0 or len(diff_keys) != 0:
+        print(f"{'Total Keys (After)'.ljust(KEY_LEN)}: {len(json_data)}")
+    # 识别有待翻译的键
+    num_miss_translation = 0
+    duplicate_items = {}
+    for key, value in json_data.items():
+        if value.startswith("#!"):
+            num_miss_translation += 1
+            if SHOW_KEYS:
+                print(f"{'Missing Translation'.ljust(KEY_LEN)}: {key}")
+        if value in duplicate_items:
+            duplicate_items[value].append(key)
+        else:
+            duplicate_items[value] = [key]
+    # 打印是否有重复的值
+    for value, keys in duplicate_items.items():
+        if len(keys) > 1:
+            print(
+                "\n".join(
+                    [f"\033[31m{'[Failed] Duplicate Value'.ljust(KEY_LEN)}: {key} -> {value}\033[0m" for key in keys]
+                )
+            )
+
+    if num_miss_translation > 0:
+        print(f"\033[31m{'[Failed] Missing Translation'.ljust(KEY_LEN)}: {num_miss_translation}\033[0m")
+    else:
+        print("\033[32m[Passed] All Keys Translated\033[0m")
+    # 将处理后的结果写入 JSON 文件
+    with open(json_file, "w", encoding="utf-8") as f:
+        json.dump(json_data, f, ensure_ascii=False, indent=4, sort_keys=SORT_KEYS)
+        f.write("\n")
+    print(f" Updated {json_file} ".center(TITLE_LEN, "=") + "\n")
+
+
+if __name__ == "__main__":
+    code_keys = scan_i18n_strings()
+    for json_file in os.listdir(I18N_JSON_DIR):
+        if json_file.endswith(r".json"):
+            json_file = os.path.join(I18N_JSON_DIR, json_file)
+            update_i18n_json(json_file, code_keys)
diff --git a/tools/my_utils.py b/tools/my_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..44d326e1ea1798968a0badf3f688ab20817cb78e
--- /dev/null
+++ b/tools/my_utils.py
@@ -0,0 +1,129 @@
+import os
+import traceback
+import ffmpeg
+import numpy as np
+import gradio as gr
+from tools.i18n.i18n import I18nAuto
+import pandas as pd
+
+i18n = I18nAuto(language=os.environ.get("language", "Auto"))
+
+
+def load_audio(file, sr):
+    try:
+        # https://github.com/openai/whisper/blob/main/whisper/audio.py#L26
+        # This launches a subprocess to decode audio while down-mixing and resampling as necessary.
+        # Requires the ffmpeg CLI and `ffmpeg-python` package to be installed.
+        file = clean_path(file)  # 防止小白拷路径头尾带了空格和"和回车
+        if os.path.exists(file) == False:
+            raise RuntimeError("You input a wrong audio path that does not exists, please fix it!")
+        out, _ = (
+            ffmpeg.input(file, threads=0)
+            .output("-", format="f32le", acodec="pcm_f32le", ac=1, ar=sr)
+            .run(cmd=["ffmpeg", "-nostdin"], capture_stdout=True, capture_stderr=True)
+        )
+    except Exception:
+        traceback.print_exc()
+        raise RuntimeError(i18n("音频加载失败"))
+
+    return np.frombuffer(out, np.float32).flatten()
+
+
+def clean_path(path_str: str):
+    if path_str.endswith(("\\", "/")):
+        return clean_path(path_str[0:-1])
+    path_str = path_str.replace("/", os.sep).replace("\\", os.sep)
+    return path_str.strip(
+        " '\n\"\u202a"
+    )  # path_str.strip(" ").strip('\'').strip("\n").strip('"').strip(" ").strip("\u202a")
+
+
+def check_for_existance(file_list: list = None, is_train=False, is_dataset_processing=False):
+    files_status = []
+    if is_train == True and file_list:
+        file_list.append(os.path.join(file_list[0], "2-name2text.txt"))
+        file_list.append(os.path.join(file_list[0], "3-bert"))
+        file_list.append(os.path.join(file_list[0], "4-cnhubert"))
+        file_list.append(os.path.join(file_list[0], "5-wav32k"))
+        file_list.append(os.path.join(file_list[0], "6-name2semantic.tsv"))
+    for file in file_list:
+        if os.path.exists(file):
+            files_status.append(True)
+        else:
+            files_status.append(False)
+    if sum(files_status) != len(files_status):
+        if is_train:
+            for file, status in zip(file_list, files_status):
+                if status:
+                    pass
+                else:
+                    gr.Warning(file)
+            gr.Warning(i18n("以下文件或文件夹不存在"))
+            return False
+        elif is_dataset_processing:
+            if files_status[0]:
+                return True
+            elif not files_status[0]:
+                gr.Warning(file_list[0])
+            elif not files_status[1] and file_list[1]:
+                gr.Warning(file_list[1])
+            gr.Warning(i18n("以下文件或文件夹不存在"))
+            return False
+        else:
+            if file_list[0]:
+                gr.Warning(file_list[0])
+                gr.Warning(i18n("以下文件或文件夹不存在"))
+            else:
+                gr.Warning(i18n("路径不能为空"))
+            return False
+    return True
+
+
+def check_details(path_list=None, is_train=False, is_dataset_processing=False):
+    if is_dataset_processing:
+        list_path, audio_path = path_list
+        if not list_path.endswith(".list"):
+            gr.Warning(i18n("请填入正确的List路径"))
+            return
+        if audio_path:
+            if not os.path.isdir(audio_path):
+                gr.Warning(i18n("请填入正确的音频文件夹路径"))
+                return
+        with open(list_path, "r", encoding="utf8") as f:
+            line = f.readline().strip("\n").split("\n")
+        wav_name, _, __, ___ = line[0].split("|")
+        wav_name = clean_path(wav_name)
+        if audio_path != "" and audio_path != None:
+            wav_name = os.path.basename(wav_name)
+            wav_path = "%s/%s" % (audio_path, wav_name)
+        else:
+            wav_path = wav_name
+        if os.path.exists(wav_path):
+            ...
+        else:
+            gr.Warning(i18n("路径错误"))
+        return
+    if is_train:
+        path_list.append(os.path.join(path_list[0], "2-name2text.txt"))
+        path_list.append(os.path.join(path_list[0], "4-cnhubert"))
+        path_list.append(os.path.join(path_list[0], "5-wav32k"))
+        path_list.append(os.path.join(path_list[0], "6-name2semantic.tsv"))
+        phone_path, hubert_path, wav_path, semantic_path = path_list[1:]
+        with open(phone_path, "r", encoding="utf-8") as f:
+            if f.read(1):
+                ...
+            else:
+                gr.Warning(i18n("缺少音素数据集"))
+        if os.listdir(hubert_path):
+            ...
+        else:
+            gr.Warning(i18n("缺少Hubert数据集"))
+        if os.listdir(wav_path):
+            ...
+        else:
+            gr.Warning(i18n("缺少音频数据集"))
+        df = pd.read_csv(semantic_path, delimiter="\t", encoding="utf-8")
+        if len(df) >= 1:
+            ...
+        else:
+            gr.Warning(i18n("缺少语义数据集"))
diff --git a/tools/slice_audio.py b/tools/slice_audio.py
new file mode 100644
index 0000000000000000000000000000000000000000..66fafa93dccd9195808a1f94ae62000466612695
--- /dev/null
+++ b/tools/slice_audio.py
@@ -0,0 +1,53 @@
+import os
+import sys
+import numpy as np
+import traceback
+from scipy.io import wavfile
+
+# parent_directory = os.path.dirname(os.path.abspath(__file__))
+# sys.path.append(parent_directory)
+from tools.my_utils import load_audio
+from slicer2 import Slicer
+
+
+def slice(inp, opt_root, threshold, min_length, min_interval, hop_size, max_sil_kept, _max, alpha, i_part, all_part):
+    os.makedirs(opt_root, exist_ok=True)
+    if os.path.isfile(inp):
+        input = [inp]
+    elif os.path.isdir(inp):
+        input = [os.path.join(inp, name) for name in sorted(list(os.listdir(inp)))]
+    else:
+        return "输入路径存在但既不是文件也不是文件夹"
+    slicer = Slicer(
+        sr=32000,  # 长音频采样率
+        threshold=int(threshold),  # 音量小于这个值视作静音的备选切割点
+        min_length=int(min_length),  # 每段最小多长，如果第一段太短一直和后面段连起来直到超过这个值
+        min_interval=int(min_interval),  # 最短切割间隔
+        hop_size=int(hop_size),  # 怎么算音量曲线，越小精度越大计算量越高（不是精度越大效果越好）
+        max_sil_kept=int(max_sil_kept),  # 切完后静音最多留多长
+    )
+    _max = float(_max)
+    alpha = float(alpha)
+    for inp_path in input[int(i_part) :: int(all_part)]:
+        # print(inp_path)
+        try:
+            name = os.path.basename(inp_path)
+            audio = load_audio(inp_path, 32000)
+            # print(audio.shape)
+            for chunk, start, end in slicer.slice(audio):  # start和end是帧数
+                tmp_max = np.abs(chunk).max()
+                if tmp_max > 1:
+                    chunk /= tmp_max
+                chunk = (chunk / tmp_max * (_max * alpha)) + (1 - alpha) * chunk
+                wavfile.write(
+                    "%s/%s_%010d_%010d.wav" % (opt_root, name, start, end),
+                    32000,
+                    # chunk.astype(np.float32),
+                    (chunk * 32767).astype(np.int16),
+                )
+        except:
+            print(inp_path, "->fail->", traceback.format_exc())
+    return "执行完毕，请检查输出文件"
+
+
+print(slice(*sys.argv[1:]))
diff --git a/tools/slicer2.py b/tools/slicer2.py
new file mode 100644
index 0000000000000000000000000000000000000000..8d80f1b4ca442e4645b5d79ddfa4f73b23f911b9
--- /dev/null
+++ b/tools/slicer2.py
@@ -0,0 +1,230 @@
+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.
+        ####音频+起始时间+终止时间
+        if len(sil_tags) == 0:
+            return [[waveform, 0, int(total_frames * self.hop_size)]]
+        else:
+            chunks = []
+            if sil_tags[0][0] > 0:
+                chunks.append([self._apply_slice(waveform, 0, sil_tags[0][0]), 0, int(sil_tags[0][0] * self.hop_size)])
+            for i in range(len(sil_tags) - 1):
+                chunks.append(
+                    [
+                        self._apply_slice(waveform, sil_tags[i][1], sil_tags[i + 1][0]),
+                        int(sil_tags[i][1] * self.hop_size),
+                        int(sil_tags[i + 1][0] * self.hop_size),
+                    ]
+                )
+            if sil_tags[-1][1] < total_frames:
+                chunks.append(
+                    [
+                        self._apply_slice(waveform, sil_tags[-1][1], total_frames),
+                        int(sil_tags[-1][1] * self.hop_size),
+                        int(total_frames * self.hop_size),
+                    ]
+                )
+            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,
+                "%s_%d.wav" % (os.path.basename(args.audio).rsplit(".", maxsplit=1)[0], i),
+            ),
+            chunk,
+            sr,
+        )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/tools/subfix_webui.py b/tools/subfix_webui.py
new file mode 100644
index 0000000000000000000000000000000000000000..85877e8fedb2fc07f1a3b9ddc019c44a807f96a3
--- /dev/null
+++ b/tools/subfix_webui.py
@@ -0,0 +1,415 @@
+import argparse
+import os
+import copy
+import json
+import uuid
+
+try:
+    import gradio.analytics as analytics
+
+    analytics.version_check = lambda: None
+except:
+    ...
+
+import librosa
+import gradio as gr
+import numpy as np
+import soundfile
+
+g_json_key_text = ""
+g_json_key_path = ""
+g_load_file = ""
+g_load_format = ""
+
+g_max_json_index = 0
+g_index = 0
+g_batch = 10
+g_text_list = []
+g_audio_list = []
+g_checkbox_list = []
+g_data_json = []
+
+
+def reload_data(index, batch):
+    global g_index
+    g_index = index
+    global g_batch
+    g_batch = batch
+    datas = g_data_json[index : index + batch]
+    output = []
+    for d in datas:
+        output.append({g_json_key_text: d[g_json_key_text], g_json_key_path: d[g_json_key_path]})
+    return output
+
+
+def b_change_index(index, batch):
+    global g_index, g_batch
+    g_index, g_batch = index, batch
+    datas = reload_data(index, batch)
+    output = []
+    for i, _ in enumerate(datas):
+        output.append(
+            # gr.Textbox(
+            #     label=f"Text {i+index}",
+            #     value=_[g_json_key_text]#text
+            # )
+            {"__type__": "update", "label": f"Text {i + index}", "value": _[g_json_key_text]}
+        )
+    for _ in range(g_batch - len(datas)):
+        output.append(
+            # gr.Textbox(
+            #     label=f"Text",
+            #     value=""
+            # )
+            {"__type__": "update", "label": "Text", "value": ""}
+        )
+    for _ in datas:
+        output.append(_[g_json_key_path])
+    for _ in range(g_batch - len(datas)):
+        output.append(None)
+    for _ in range(g_batch):
+        output.append(False)
+    return output
+
+
+def b_next_index(index, batch):
+    b_save_file()
+    if (index + batch) <= g_max_json_index:
+        return index + batch, *b_change_index(index + batch, batch)
+    else:
+        return index, *b_change_index(index, batch)
+
+
+def b_previous_index(index, batch):
+    b_save_file()
+    if (index - batch) >= 0:
+        return index - batch, *b_change_index(index - batch, batch)
+    else:
+        return 0, *b_change_index(0, batch)
+
+
+def b_submit_change(*text_list):
+    global g_data_json
+    change = False
+    for i, new_text in enumerate(text_list):
+        if g_index + i <= g_max_json_index:
+            new_text = new_text.strip() + " "
+            if g_data_json[g_index + i][g_json_key_text] != new_text:
+                g_data_json[g_index + i][g_json_key_text] = new_text
+                change = True
+    if change:
+        b_save_file()
+    return g_index, *b_change_index(g_index, g_batch)
+
+
+def b_delete_audio(*checkbox_list):
+    global g_data_json, g_index, g_max_json_index
+    b_save_file()
+    change = False
+    for i, checkbox in reversed(list(enumerate(checkbox_list))):
+        if g_index + i < len(g_data_json):
+            if checkbox == True:
+                g_data_json.pop(g_index + i)
+                change = True
+
+    g_max_json_index = len(g_data_json) - 1
+    if g_index > g_max_json_index:
+        g_index = g_max_json_index
+        g_index = g_index if g_index >= 0 else 0
+    if change:
+        b_save_file()
+    # return gr.Slider(value=g_index, maximum=(g_max_json_index if g_max_json_index>=0 else 0)), *b_change_index(g_index, g_batch)
+    return {
+        "value": g_index,
+        "__type__": "update",
+        "maximum": (g_max_json_index if g_max_json_index >= 0 else 0),
+    }, *b_change_index(g_index, g_batch)
+
+
+def b_invert_selection(*checkbox_list):
+    new_list = [not item if item is True else True for item in checkbox_list]
+    return new_list
+
+
+def get_next_path(filename):
+    base_dir = os.path.dirname(filename)
+    base_name = os.path.splitext(os.path.basename(filename))[0]
+    for i in range(100):
+        new_path = os.path.join(base_dir, f"{base_name}_{str(i).zfill(2)}.wav")
+        if not os.path.exists(new_path):
+            return new_path
+    return os.path.join(base_dir, f"{str(uuid.uuid4())}.wav")
+
+
+def b_audio_split(audio_breakpoint, *checkbox_list):
+    global g_data_json, g_max_json_index
+    checked_index = []
+    for i, checkbox in enumerate(checkbox_list):
+        if checkbox == True and g_index + i < len(g_data_json):
+            checked_index.append(g_index + i)
+    if len(checked_index) == 1:
+        index = checked_index[0]
+        audio_json = copy.deepcopy(g_data_json[index])
+        path = audio_json[g_json_key_path]
+        data, sample_rate = librosa.load(path, sr=None, mono=True)
+        audio_maxframe = len(data)
+        break_frame = int(audio_breakpoint * sample_rate)
+
+        if break_frame >= 1 and break_frame < audio_maxframe:
+            audio_first = data[0:break_frame]
+            audio_second = data[break_frame:]
+            nextpath = get_next_path(path)
+            soundfile.write(nextpath, audio_second, sample_rate)
+            soundfile.write(path, audio_first, sample_rate)
+            g_data_json.insert(index + 1, audio_json)
+            g_data_json[index + 1][g_json_key_path] = nextpath
+            b_save_file()
+
+    g_max_json_index = len(g_data_json) - 1
+    # return gr.Slider(value=g_index, maximum=g_max_json_index), *b_change_index(g_index, g_batch)
+    return {"value": g_index, "maximum": g_max_json_index, "__type__": "update"}, *b_change_index(g_index, g_batch)
+
+
+def b_merge_audio(interval_r, *checkbox_list):
+    global g_data_json, g_max_json_index
+    b_save_file()
+    checked_index = []
+    audios_path = []
+    audios_text = []
+    for i, checkbox in enumerate(checkbox_list):
+        if checkbox == True and g_index + i < len(g_data_json):
+            checked_index.append(g_index + i)
+
+    if len(checked_index) > 1:
+        for i in checked_index:
+            audios_path.append(g_data_json[i][g_json_key_path])
+            audios_text.append(g_data_json[i][g_json_key_text])
+        for i in reversed(checked_index[1:]):
+            g_data_json.pop(i)
+
+        base_index = checked_index[0]
+        base_path = audios_path[0]
+        g_data_json[base_index][g_json_key_text] = "".join(audios_text)
+
+        audio_list = []
+        l_sample_rate = None
+        for i, path in enumerate(audios_path):
+            data, sample_rate = librosa.load(path, sr=l_sample_rate, mono=True)
+            l_sample_rate = sample_rate
+            if i > 0:
+                silence = np.zeros(int(l_sample_rate * interval_r))
+                audio_list.append(silence)
+
+            audio_list.append(data)
+
+        audio_concat = np.concatenate(audio_list)
+
+        soundfile.write(base_path, audio_concat, l_sample_rate)
+
+        b_save_file()
+
+    g_max_json_index = len(g_data_json) - 1
+
+    # return gr.Slider(value=g_index, maximum=g_max_json_index), *b_change_index(g_index, g_batch)
+    return {"value": g_index, "maximum": g_max_json_index, "__type__": "update"}, *b_change_index(g_index, g_batch)
+
+
+def b_save_json():
+    with open(g_load_file, "w", encoding="utf-8") as file:
+        for data in g_data_json:
+            file.write(f"{json.dumps(data, ensure_ascii=False)}\n")
+
+
+def b_save_list():
+    with open(g_load_file, "w", encoding="utf-8") as file:
+        for data in g_data_json:
+            wav_path = data["wav_path"]
+            speaker_name = data["speaker_name"]
+            language = data["language"]
+            text = data["text"]
+            file.write(f"{wav_path}|{speaker_name}|{language}|{text}".strip() + "\n")
+
+
+def b_load_json():
+    global g_data_json, g_max_json_index
+    with open(g_load_file, "r", encoding="utf-8") as file:
+        g_data_json = file.readlines()
+        g_data_json = [json.loads(line) for line in g_data_json]
+        g_max_json_index = len(g_data_json) - 1
+
+
+def b_load_list():
+    global g_data_json, g_max_json_index
+    with open(g_load_file, "r", encoding="utf-8") as source:
+        data_list = source.readlines()
+        for _ in data_list:
+            data = _.split("|")
+            if len(data) == 4:
+                wav_path, speaker_name, language, text = data
+                g_data_json.append(
+                    {"wav_path": wav_path, "speaker_name": speaker_name, "language": language, "text": text.strip()}
+                )
+            else:
+                print("error line:", data)
+        g_max_json_index = len(g_data_json) - 1
+
+
+def b_save_file():
+    if g_load_format == "json":
+        b_save_json()
+    elif g_load_format == "list":
+        b_save_list()
+
+
+def b_load_file():
+    if g_load_format == "json":
+        b_load_json()
+    elif g_load_format == "list":
+        b_load_list()
+
+
+def set_global(load_json, load_list, json_key_text, json_key_path, batch):
+    global g_json_key_text, g_json_key_path, g_load_file, g_load_format, g_batch
+
+    g_batch = int(batch)
+
+    if load_json != "None":
+        g_load_format = "json"
+        g_load_file = load_json
+    elif load_list != "None":
+        g_load_format = "list"
+        g_load_file = load_list
+    else:
+        g_load_format = "list"
+        g_load_file = "demo.list"
+
+    g_json_key_text = json_key_text
+    g_json_key_path = json_key_path
+
+    b_load_file()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Process some integers.")
+    parser.add_argument("--load_json", default="None", help="source file, like demo.json")
+    parser.add_argument("--is_share", default="False", help="whether webui is_share=True")
+    parser.add_argument("--load_list", default="None", help="source file, like demo.list")
+    parser.add_argument("--webui_port_subfix", default=9871, help="source file, like demo.list")
+    parser.add_argument("--json_key_text", default="text", help="the text key name in json, Default: text")
+    parser.add_argument("--json_key_path", default="wav_path", help="the path key name in json, Default: wav_path")
+    parser.add_argument("--g_batch", default=10, help="max number g_batch wav to display, Default: 10")
+
+    args = parser.parse_args()
+
+    set_global(args.load_json, args.load_list, args.json_key_text, args.json_key_path, args.g_batch)
+
+    with gr.Blocks() as demo:
+        with gr.Row():
+            btn_change_index = gr.Button("Change Index")
+            btn_submit_change = gr.Button("Submit Text")
+            btn_merge_audio = gr.Button("Merge Audio")
+            btn_delete_audio = gr.Button("Delete Audio")
+            btn_previous_index = gr.Button("Previous Index")
+            btn_next_index = gr.Button("Next Index")
+
+        with gr.Row():
+            index_slider = gr.Slider(minimum=0, maximum=g_max_json_index, value=g_index, step=1, label="Index", scale=3)
+            splitpoint_slider = gr.Slider(
+                minimum=0, maximum=120.0, value=0, step=0.1, label="Audio Split Point(s)", scale=3
+            )
+            btn_audio_split = gr.Button("Split Audio", scale=1)
+            btn_save_json = gr.Button("Save File", visible=True, scale=1)
+            btn_invert_selection = gr.Button("Invert Selection", scale=1)
+
+        with gr.Row():
+            with gr.Column():
+                for _ in range(0, g_batch):
+                    with gr.Row():
+                        text = gr.Textbox(label="Text", visible=True, scale=5)
+                        audio_output = gr.Audio(label="Output Audio", visible=True, scale=5)
+                        audio_check = gr.Checkbox(label="Yes", show_label=True, info="Choose Audio", scale=1)
+                        g_text_list.append(text)
+                        g_audio_list.append(audio_output)
+                        g_checkbox_list.append(audio_check)
+
+        with gr.Row():
+            batchsize_slider = gr.Slider(
+                minimum=1, maximum=g_batch, value=g_batch, step=1, label="Batch Size", scale=3, interactive=False
+            )
+            interval_slider = gr.Slider(minimum=0, maximum=2, value=0, step=0.01, label="Interval", scale=3)
+            btn_theme_dark = gr.Button("Light Theme", link="?__theme=light", scale=1)
+            btn_theme_light = gr.Button("Dark Theme", link="?__theme=dark", scale=1)
+
+        btn_change_index.click(
+            b_change_index,
+            inputs=[
+                index_slider,
+                batchsize_slider,
+            ],
+            outputs=[*g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_submit_change.click(
+            b_submit_change,
+            inputs=[
+                *g_text_list,
+            ],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_previous_index.click(
+            b_previous_index,
+            inputs=[
+                index_slider,
+                batchsize_slider,
+            ],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_next_index.click(
+            b_next_index,
+            inputs=[
+                index_slider,
+                batchsize_slider,
+            ],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_delete_audio.click(
+            b_delete_audio,
+            inputs=[*g_checkbox_list],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_merge_audio.click(
+            b_merge_audio,
+            inputs=[interval_slider, *g_checkbox_list],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_audio_split.click(
+            b_audio_split,
+            inputs=[splitpoint_slider, *g_checkbox_list],
+            outputs=[index_slider, *g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+        btn_invert_selection.click(b_invert_selection, inputs=[*g_checkbox_list], outputs=[*g_checkbox_list])
+
+        btn_save_json.click(b_save_file)
+
+        demo.load(
+            b_change_index,
+            inputs=[
+                index_slider,
+                batchsize_slider,
+            ],
+            outputs=[*g_text_list, *g_audio_list, *g_checkbox_list],
+        )
+
+    demo.launch(
+        server_name="0.0.0.0",
+        inbrowser=True,
+        quiet=True,
+        share=eval(args.is_share),
+        server_port=int(args.webui_port_subfix),
+    )
diff --git a/tools/uvr5/bs_roformer/__init__.py b/tools/uvr5/bs_roformer/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/tools/uvr5/bs_roformer/attend.py b/tools/uvr5/bs_roformer/attend.py
new file mode 100644
index 0000000000000000000000000000000000000000..29dad18e1b05a1926f3dd142ddec1dd6401e8243
--- /dev/null
+++ b/tools/uvr5/bs_roformer/attend.py
@@ -0,0 +1,67 @@
+from packaging import version
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+class Attend(nn.Module):
+    def __init__(self, dropout=0.0, flash=False, scale=None):
+        super().__init__()
+        self.scale = scale
+        self.dropout = dropout
+        self.attn_dropout = nn.Dropout(dropout)
+
+        self.flash = flash
+        assert not (flash and version.parse(torch.__version__) < version.parse("2.0.0")), (
+            "in order to use flash attention, you must be using pytorch 2.0 or above"
+        )
+
+    def flash_attn(self, q, k, v):
+        # _, heads, q_len, _, k_len, is_cuda, device = *q.shape, k.shape[-2], q.is_cuda, q.device
+
+        if exists(self.scale):
+            default_scale = q.shape[-1] ** -0.5
+            q = q * (self.scale / default_scale)
+
+        # pytorch 2.0 flash attn: q, k, v, mask, dropout, softmax_scale
+        # with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True):
+        return F.scaled_dot_product_attention(q, k, v, dropout_p=self.dropout if self.training else 0.0)
+
+    def forward(self, q, k, v):
+        """
+        einstein notation
+        b - batch
+        h - heads
+        n, i, j - sequence length (base sequence length, source, target)
+        d - feature dimension
+        """
+
+        # q_len, k_len, device = q.shape[-2], k.shape[-2], q.device
+
+        scale = default(self.scale, q.shape[-1] ** -0.5)
+
+        if self.flash:
+            return self.flash_attn(q, k, v)
+
+        # similarity
+
+        sim = einsum("b h i d, b h j d -> b h i j", q, k) * scale
+
+        # attention
+
+        attn = sim.softmax(dim=-1)
+        attn = self.attn_dropout(attn)
+
+        # aggregate values
+
+        out = einsum("b h i j, b h j d -> b h i d", attn, v)
+
+        return out
diff --git a/tools/uvr5/bs_roformer/bs_roformer.py b/tools/uvr5/bs_roformer/bs_roformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..8c9d1f300435f26b6a11f13b117f88e1b308aeb0
--- /dev/null
+++ b/tools/uvr5/bs_roformer/bs_roformer.py
@@ -0,0 +1,626 @@
+from functools import partial
+
+import torch
+from torch import nn
+from torch.nn import Module, ModuleList
+import torch.nn.functional as F
+
+from bs_roformer.attend import Attend
+from torch.utils.checkpoint import checkpoint
+
+from typing import Tuple, Optional, Callable
+# from beartype.typing import Tuple, Optional, List, Callable
+# from beartype import beartype
+
+from rotary_embedding_torch import RotaryEmbedding
+
+from einops import rearrange, pack, unpack
+from einops.layers.torch import Rearrange
+
+# helper functions
+
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
+
+# norm
+
+
+def l2norm(t):
+    return F.normalize(t, dim=-1, p=2)
+
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+
+class FeedForward(Module):
+    def __init__(self, dim, mult=4, dropout=0.0):
+        super().__init__()
+        dim_inner = int(dim * mult)
+        self.net = nn.Sequential(
+            RMSNorm(dim),
+            nn.Linear(dim, dim_inner),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_inner, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0, rotary_embed=None, flash=True):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head**-0.5
+        dim_inner = heads * dim_head
+
+        self.rotary_embed = rotary_embed
+
+        self.attend = Attend(flash=flash, dropout=dropout)
+
+        self.norm = RMSNorm(dim)
+        self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
+
+        self.to_gates = nn.Linear(dim, heads)
+
+        self.to_out = nn.Sequential(nn.Linear(dim_inner, dim, bias=False), nn.Dropout(dropout))
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        q, k, v = rearrange(self.to_qkv(x), "b n (qkv h d) -> qkv b h n d", qkv=3, h=self.heads)
+
+        if exists(self.rotary_embed):
+            q = self.rotary_embed.rotate_queries_or_keys(q)
+            k = self.rotary_embed.rotate_queries_or_keys(k)
+
+        out = self.attend(q, k, v)
+
+        gates = self.to_gates(x)
+        out = out * rearrange(gates, "b n h -> b h n 1").sigmoid()
+
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class LinearAttention(Module):
+    """
+    this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
+    """
+
+    # @beartype
+    def __init__(self, *, dim, dim_head=32, heads=8, scale=8, flash=False, dropout=0.0):
+        super().__init__()
+        dim_inner = dim_head * heads
+        self.norm = RMSNorm(dim)
+
+        self.to_qkv = nn.Sequential(
+            nn.Linear(dim, dim_inner * 3, bias=False), Rearrange("b n (qkv h d) -> qkv b h d n", qkv=3, h=heads)
+        )
+
+        self.temperature = nn.Parameter(torch.ones(heads, 1, 1))
+
+        self.attend = Attend(scale=scale, dropout=dropout, flash=flash)
+
+        self.to_out = nn.Sequential(Rearrange("b h d n -> b n (h d)"), nn.Linear(dim_inner, dim, bias=False))
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x)
+
+        q, k = map(l2norm, (q, k))
+        q = q * self.temperature.exp()
+
+        out = self.attend(q, k, v)
+
+        return self.to_out(out)
+
+
+class Transformer(Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        depth,
+        dim_head=64,
+        heads=8,
+        attn_dropout=0.0,
+        ff_dropout=0.0,
+        ff_mult=4,
+        norm_output=True,
+        rotary_embed=None,
+        flash_attn=True,
+        linear_attn=False,
+    ):
+        super().__init__()
+        self.layers = ModuleList([])
+
+        for _ in range(depth):
+            if linear_attn:
+                attn = LinearAttention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout, flash=flash_attn)
+            else:
+                attn = Attention(
+                    dim=dim,
+                    dim_head=dim_head,
+                    heads=heads,
+                    dropout=attn_dropout,
+                    rotary_embed=rotary_embed,
+                    flash=flash_attn,
+                )
+
+            self.layers.append(ModuleList([attn, FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)]))
+
+        self.norm = RMSNorm(dim) if norm_output else nn.Identity()
+
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+
+        return self.norm(x)
+
+
+# bandsplit module
+
+
+class BandSplit(Module):
+    # @beartype
+    def __init__(self, dim, dim_inputs: Tuple[int, ...]):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_features = ModuleList([])
+
+        for dim_in in dim_inputs:
+            net = nn.Sequential(RMSNorm(dim_in), nn.Linear(dim_in, dim))
+
+            self.to_features.append(net)
+
+    def forward(self, x):
+        x = x.split(self.dim_inputs, dim=-1)
+
+        outs = []
+        for split_input, to_feature in zip(x, self.to_features):
+            split_output = to_feature(split_input)
+            outs.append(split_output)
+
+        return torch.stack(outs, dim=-2)
+
+
+def MLP(dim_in, dim_out, dim_hidden=None, depth=1, activation=nn.Tanh):
+    dim_hidden = default(dim_hidden, dim_in)
+
+    net = []
+    dims = (dim_in, *((dim_hidden,) * (depth - 1)), dim_out)
+
+    for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
+        is_last = ind == (len(dims) - 2)
+
+        net.append(nn.Linear(layer_dim_in, layer_dim_out))
+
+        if is_last:
+            continue
+
+        net.append(activation())
+
+    return nn.Sequential(*net)
+
+
+class MaskEstimator(Module):
+    # @beartype
+    def __init__(self, dim, dim_inputs: Tuple[int, ...], depth, mlp_expansion_factor=4):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_freqs = ModuleList([])
+        dim_hidden = dim * mlp_expansion_factor
+
+        for dim_in in dim_inputs:
+            net = []
+
+            mlp = nn.Sequential(MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth), nn.GLU(dim=-1))
+
+            self.to_freqs.append(mlp)
+
+    def forward(self, x):
+        x = x.unbind(dim=-2)
+
+        outs = []
+
+        for band_features, mlp in zip(x, self.to_freqs):
+            freq_out = mlp(band_features)
+            outs.append(freq_out)
+
+        return torch.cat(outs, dim=-1)
+
+
+# main class
+
+DEFAULT_FREQS_PER_BANDS = (
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    4,
+    12,
+    12,
+    12,
+    12,
+    12,
+    12,
+    12,
+    12,
+    24,
+    24,
+    24,
+    24,
+    24,
+    24,
+    24,
+    24,
+    48,
+    48,
+    48,
+    48,
+    48,
+    48,
+    48,
+    48,
+    128,
+    129,
+)
+
+
+class BSRoformer(Module):
+    # @beartype
+    def __init__(
+        self,
+        dim,
+        *,
+        depth,
+        stereo=False,
+        num_stems=1,
+        time_transformer_depth=2,
+        freq_transformer_depth=2,
+        linear_transformer_depth=0,
+        freqs_per_bands: Tuple[int, ...] = DEFAULT_FREQS_PER_BANDS,
+        # in the paper, they divide into ~60 bands, test with 1 for starters
+        dim_head=64,
+        heads=8,
+        attn_dropout=0.0,
+        ff_dropout=0.0,
+        flash_attn=True,
+        dim_freqs_in=1025,
+        stft_n_fft=2048,
+        stft_hop_length=512,
+        # 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
+        stft_win_length=2048,
+        stft_normalized=False,
+        stft_window_fn: Optional[Callable] = None,
+        mask_estimator_depth=2,
+        multi_stft_resolution_loss_weight=1.0,
+        multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
+        multi_stft_hop_size=147,
+        multi_stft_normalized=False,
+        multi_stft_window_fn: Callable = torch.hann_window,
+        mlp_expansion_factor=4,
+        use_torch_checkpoint=False,
+        skip_connection=False,
+    ):
+        super().__init__()
+
+        self.stereo = stereo
+        self.audio_channels = 2 if stereo else 1
+        self.num_stems = num_stems
+        self.use_torch_checkpoint = use_torch_checkpoint
+        self.skip_connection = skip_connection
+
+        self.layers = ModuleList([])
+
+        transformer_kwargs = dict(
+            dim=dim,
+            heads=heads,
+            dim_head=dim_head,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            flash_attn=flash_attn,
+            norm_output=False,
+        )
+
+        time_rotary_embed = RotaryEmbedding(dim=dim_head)
+        freq_rotary_embed = RotaryEmbedding(dim=dim_head)
+
+        for _ in range(depth):
+            tran_modules = []
+            if linear_transformer_depth > 0:
+                tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
+            tran_modules.append(
+                Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, **transformer_kwargs)
+            )
+            tran_modules.append(
+                Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, **transformer_kwargs)
+            )
+            self.layers.append(nn.ModuleList(tran_modules))
+
+        self.final_norm = RMSNorm(dim)
+
+        self.stft_kwargs = dict(
+            n_fft=stft_n_fft, hop_length=stft_hop_length, win_length=stft_win_length, normalized=stft_normalized
+        )
+
+        self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
+
+        freqs = torch.stft(
+            torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_win_length), return_complex=True
+        ).shape[1]
+
+        assert len(freqs_per_bands) > 1
+        assert sum(freqs_per_bands) == freqs, (
+            f"the number of freqs in the bands must equal {freqs} based on the STFT settings, but got {sum(freqs_per_bands)}"
+        )
+
+        freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in freqs_per_bands)
+
+        self.band_split = BandSplit(dim=dim, dim_inputs=freqs_per_bands_with_complex)
+
+        self.mask_estimators = nn.ModuleList([])
+
+        for _ in range(num_stems):
+            mask_estimator = MaskEstimator(
+                dim=dim,
+                dim_inputs=freqs_per_bands_with_complex,
+                depth=mask_estimator_depth,
+                mlp_expansion_factor=mlp_expansion_factor,
+            )
+
+            self.mask_estimators.append(mask_estimator)
+
+        # for the multi-resolution stft loss
+
+        self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
+        self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
+        self.multi_stft_n_fft = stft_n_fft
+        self.multi_stft_window_fn = multi_stft_window_fn
+
+        self.multi_stft_kwargs = dict(hop_length=multi_stft_hop_size, normalized=multi_stft_normalized)
+
+    def forward(self, raw_audio, target=None, return_loss_breakdown=False):
+        """
+        einops
+
+        b - batch
+        f - freq
+        t - time
+        s - audio channel (1 for mono, 2 for stereo)
+        n - number of 'stems'
+        c - complex (2)
+        d - feature dimension
+        """
+
+        device = raw_audio.device
+
+        # defining whether model is loaded on MPS (MacOS GPU accelerator)
+        x_is_mps = True if device.type == "mps" else False
+
+        if raw_audio.ndim == 2:
+            raw_audio = rearrange(raw_audio, "b t -> b 1 t")
+
+        channels = raw_audio.shape[1]
+        assert (not self.stereo and channels == 1) or (self.stereo and channels == 2), (
+            "stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)"
+        )
+
+        # to stft
+
+        raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, "* t")
+
+        stft_window = self.stft_window_fn(device=device)
+
+        # RuntimeError: FFT operations are only supported on MacOS 14+
+        # Since it's tedious to define whether we're on correct MacOS version - simple try-catch is used
+        try:
+            stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
+        except:
+            stft_repr = torch.stft(
+                raw_audio.cpu() if x_is_mps else raw_audio,
+                **self.stft_kwargs,
+                window=stft_window.cpu() if x_is_mps else stft_window,
+                return_complex=True,
+            ).to(device)
+
+        stft_repr = torch.view_as_real(stft_repr)
+
+        stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, "* f t c")
+
+        # merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
+        stft_repr = rearrange(stft_repr, "b s f t c -> b (f s) t c")
+
+        x = rearrange(stft_repr, "b f t c -> b t (f c)")
+
+        if self.use_torch_checkpoint:
+            x = checkpoint(self.band_split, x, use_reentrant=False)
+        else:
+            x = self.band_split(x)
+
+        # axial / hierarchical attention
+
+        store = [None] * len(self.layers)
+        for i, transformer_block in enumerate(self.layers):
+            if len(transformer_block) == 3:
+                linear_transformer, time_transformer, freq_transformer = transformer_block
+
+                x, ft_ps = pack([x], "b * d")
+                if self.use_torch_checkpoint:
+                    x = checkpoint(linear_transformer, x, use_reentrant=False)
+                else:
+                    x = linear_transformer(x)
+                (x,) = unpack(x, ft_ps, "b * d")
+            else:
+                time_transformer, freq_transformer = transformer_block
+
+            if self.skip_connection:
+                # Sum all previous
+                for j in range(i):
+                    x = x + store[j]
+
+            x = rearrange(x, "b t f d -> b f t d")
+            x, ps = pack([x], "* t d")
+
+            if self.use_torch_checkpoint:
+                x = checkpoint(time_transformer, x, use_reentrant=False)
+            else:
+                x = time_transformer(x)
+
+            (x,) = unpack(x, ps, "* t d")
+            x = rearrange(x, "b f t d -> b t f d")
+            x, ps = pack([x], "* f d")
+
+            if self.use_torch_checkpoint:
+                x = checkpoint(freq_transformer, x, use_reentrant=False)
+            else:
+                x = freq_transformer(x)
+
+            (x,) = unpack(x, ps, "* f d")
+
+            if self.skip_connection:
+                store[i] = x
+
+        x = self.final_norm(x)
+
+        num_stems = len(self.mask_estimators)
+
+        if self.use_torch_checkpoint:
+            mask = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
+        else:
+            mask = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
+        mask = rearrange(mask, "b n t (f c) -> b n f t c", c=2)
+
+        # modulate frequency representation
+
+        stft_repr = rearrange(stft_repr, "b f t c -> b 1 f t c")
+
+        # complex number multiplication
+
+        stft_repr = torch.view_as_complex(stft_repr)
+        mask = torch.view_as_complex(mask)
+
+        stft_repr = stft_repr * mask
+
+        # istft
+
+        stft_repr = rearrange(stft_repr, "b n (f s) t -> (b n s) f t", s=self.audio_channels)
+
+        # same as torch.stft() fix for MacOS MPS above
+        try:
+            recon_audio = torch.istft(
+                stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False, length=raw_audio.shape[-1]
+            )
+        except:
+            recon_audio = torch.istft(
+                stft_repr.cpu() if x_is_mps else stft_repr,
+                **self.stft_kwargs,
+                window=stft_window.cpu() if x_is_mps else stft_window,
+                return_complex=False,
+                length=raw_audio.shape[-1],
+            ).to(device)
+
+        recon_audio = rearrange(recon_audio, "(b n s) t -> b n s t", s=self.audio_channels, n=num_stems)
+
+        if num_stems == 1:
+            recon_audio = rearrange(recon_audio, "b 1 s t -> b s t")
+
+        # if a target is passed in, calculate loss for learning
+
+        if not exists(target):
+            return recon_audio
+
+        if self.num_stems > 1:
+            assert target.ndim == 4 and target.shape[1] == self.num_stems
+
+        if target.ndim == 2:
+            target = rearrange(target, "... t -> ... 1 t")
+
+        target = target[..., : recon_audio.shape[-1]]  # protect against lost length on istft
+
+        loss = F.l1_loss(recon_audio, target)
+
+        multi_stft_resolution_loss = 0.0
+
+        for window_size in self.multi_stft_resolutions_window_sizes:
+            res_stft_kwargs = dict(
+                n_fft=max(window_size, self.multi_stft_n_fft),  # not sure what n_fft is across multi resolution stft
+                win_length=window_size,
+                return_complex=True,
+                window=self.multi_stft_window_fn(window_size, device=device),
+                **self.multi_stft_kwargs,
+            )
+
+            recon_Y = torch.stft(rearrange(recon_audio, "... s t -> (... s) t"), **res_stft_kwargs)
+            target_Y = torch.stft(rearrange(target, "... s t -> (... s) t"), **res_stft_kwargs)
+
+            multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
+
+        weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
+
+        total_loss = loss + weighted_multi_resolution_loss
+
+        if not return_loss_breakdown:
+            return total_loss
+
+        return total_loss, (loss, multi_stft_resolution_loss)
diff --git a/tools/uvr5/bs_roformer/mel_band_roformer.py b/tools/uvr5/bs_roformer/mel_band_roformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..78a390400133307c2a41bf044abf39d165180370
--- /dev/null
+++ b/tools/uvr5/bs_roformer/mel_band_roformer.py
@@ -0,0 +1,606 @@
+from functools import partial
+
+import torch
+from torch import nn
+from torch.nn import Module, ModuleList
+import torch.nn.functional as F
+
+from bs_roformer.attend import Attend
+from torch.utils.checkpoint import checkpoint
+
+from typing import Tuple, Optional, Callable
+# from beartype.typing import Tuple, Optional, List, Callable
+# from beartype import beartype
+
+from rotary_embedding_torch import RotaryEmbedding
+
+from einops import rearrange, pack, unpack, reduce, repeat
+from einops.layers.torch import Rearrange
+
+from librosa import filters
+
+
+# helper functions
+
+
+def exists(val):
+    return val is not None
+
+
+def default(v, d):
+    return v if exists(v) else d
+
+
+def pack_one(t, pattern):
+    return pack([t], pattern)
+
+
+def unpack_one(t, ps, pattern):
+    return unpack(t, ps, pattern)[0]
+
+
+def pad_at_dim(t, pad, dim=-1, value=0.0):
+    dims_from_right = (-dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = (0, 0) * dims_from_right
+    return F.pad(t, (*zeros, *pad), value=value)
+
+
+def l2norm(t):
+    return F.normalize(t, dim=-1, p=2)
+
+
+# norm
+
+
+class RMSNorm(Module):
+    def __init__(self, dim):
+        super().__init__()
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        return F.normalize(x, dim=-1) * self.scale * self.gamma
+
+
+# attention
+
+
+class FeedForward(Module):
+    def __init__(self, dim, mult=4, dropout=0.0):
+        super().__init__()
+        dim_inner = int(dim * mult)
+        self.net = nn.Sequential(
+            RMSNorm(dim),
+            nn.Linear(dim, dim_inner),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(dim_inner, dim),
+            nn.Dropout(dropout),
+        )
+
+    def forward(self, x):
+        return self.net(x)
+
+
+class Attention(Module):
+    def __init__(self, dim, heads=8, dim_head=64, dropout=0.0, rotary_embed=None, flash=True):
+        super().__init__()
+        self.heads = heads
+        self.scale = dim_head**-0.5
+        dim_inner = heads * dim_head
+
+        self.rotary_embed = rotary_embed
+
+        self.attend = Attend(flash=flash, dropout=dropout)
+
+        self.norm = RMSNorm(dim)
+        self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
+
+        self.to_gates = nn.Linear(dim, heads)
+
+        self.to_out = nn.Sequential(nn.Linear(dim_inner, dim, bias=False), nn.Dropout(dropout))
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        q, k, v = rearrange(self.to_qkv(x), "b n (qkv h d) -> qkv b h n d", qkv=3, h=self.heads)
+
+        if exists(self.rotary_embed):
+            q = self.rotary_embed.rotate_queries_or_keys(q)
+            k = self.rotary_embed.rotate_queries_or_keys(k)
+
+        out = self.attend(q, k, v)
+
+        gates = self.to_gates(x)
+        out = out * rearrange(gates, "b n h -> b h n 1").sigmoid()
+
+        out = rearrange(out, "b h n d -> b n (h d)")
+        return self.to_out(out)
+
+
+class LinearAttention(Module):
+    """
+    this flavor of linear attention proposed in https://arxiv.org/abs/2106.09681 by El-Nouby et al.
+    """
+
+    # @beartype
+    def __init__(self, *, dim, dim_head=32, heads=8, scale=8, flash=False, dropout=0.0):
+        super().__init__()
+        dim_inner = dim_head * heads
+        self.norm = RMSNorm(dim)
+
+        self.to_qkv = nn.Sequential(
+            nn.Linear(dim, dim_inner * 3, bias=False), Rearrange("b n (qkv h d) -> qkv b h d n", qkv=3, h=heads)
+        )
+
+        self.temperature = nn.Parameter(torch.ones(heads, 1, 1))
+
+        self.attend = Attend(scale=scale, dropout=dropout, flash=flash)
+
+        self.to_out = nn.Sequential(Rearrange("b h d n -> b n (h d)"), nn.Linear(dim_inner, dim, bias=False))
+
+    def forward(self, x):
+        x = self.norm(x)
+
+        q, k, v = self.to_qkv(x)
+
+        q, k = map(l2norm, (q, k))
+        q = q * self.temperature.exp()
+
+        out = self.attend(q, k, v)
+
+        return self.to_out(out)
+
+
+class Transformer(Module):
+    def __init__(
+        self,
+        *,
+        dim,
+        depth,
+        dim_head=64,
+        heads=8,
+        attn_dropout=0.0,
+        ff_dropout=0.0,
+        ff_mult=4,
+        norm_output=True,
+        rotary_embed=None,
+        flash_attn=True,
+        linear_attn=False,
+    ):
+        super().__init__()
+        self.layers = ModuleList([])
+
+        for _ in range(depth):
+            if linear_attn:
+                attn = LinearAttention(dim=dim, dim_head=dim_head, heads=heads, dropout=attn_dropout, flash=flash_attn)
+            else:
+                attn = Attention(
+                    dim=dim,
+                    dim_head=dim_head,
+                    heads=heads,
+                    dropout=attn_dropout,
+                    rotary_embed=rotary_embed,
+                    flash=flash_attn,
+                )
+
+            self.layers.append(ModuleList([attn, FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)]))
+
+        self.norm = RMSNorm(dim) if norm_output else nn.Identity()
+
+    def forward(self, x):
+        for attn, ff in self.layers:
+            x = attn(x) + x
+            x = ff(x) + x
+
+        return self.norm(x)
+
+
+# bandsplit module
+
+
+class BandSplit(Module):
+    # @beartype
+    def __init__(self, dim, dim_inputs: Tuple[int, ...]):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_features = ModuleList([])
+
+        for dim_in in dim_inputs:
+            net = nn.Sequential(RMSNorm(dim_in), nn.Linear(dim_in, dim))
+
+            self.to_features.append(net)
+
+    def forward(self, x):
+        x = x.split(self.dim_inputs, dim=-1)
+
+        outs = []
+        for split_input, to_feature in zip(x, self.to_features):
+            split_output = to_feature(split_input)
+            outs.append(split_output)
+
+        return torch.stack(outs, dim=-2)
+
+
+def MLP(dim_in, dim_out, dim_hidden=None, depth=1, activation=nn.Tanh):
+    dim_hidden = default(dim_hidden, dim_in)
+
+    net = []
+    dims = (dim_in, *((dim_hidden,) * depth), dim_out)
+
+    for ind, (layer_dim_in, layer_dim_out) in enumerate(zip(dims[:-1], dims[1:])):
+        is_last = ind == (len(dims) - 2)
+
+        net.append(nn.Linear(layer_dim_in, layer_dim_out))
+
+        if is_last:
+            continue
+
+        net.append(activation())
+
+    return nn.Sequential(*net)
+
+
+class MaskEstimator(Module):
+    # @beartype
+    def __init__(self, dim, dim_inputs: Tuple[int, ...], depth, mlp_expansion_factor=4):
+        super().__init__()
+        self.dim_inputs = dim_inputs
+        self.to_freqs = ModuleList([])
+        dim_hidden = dim * mlp_expansion_factor
+
+        for dim_in in dim_inputs:
+            net = []
+
+            mlp = nn.Sequential(MLP(dim, dim_in * 2, dim_hidden=dim_hidden, depth=depth), nn.GLU(dim=-1))
+
+            self.to_freqs.append(mlp)
+
+    def forward(self, x):
+        x = x.unbind(dim=-2)
+
+        outs = []
+
+        for band_features, mlp in zip(x, self.to_freqs):
+            freq_out = mlp(band_features)
+            outs.append(freq_out)
+
+        return torch.cat(outs, dim=-1)
+
+
+# main class
+
+
+class MelBandRoformer(Module):
+    # @beartype
+    def __init__(
+        self,
+        dim,
+        *,
+        depth,
+        stereo=False,
+        num_stems=1,
+        time_transformer_depth=2,
+        freq_transformer_depth=2,
+        linear_transformer_depth=0,
+        num_bands=60,
+        dim_head=64,
+        heads=8,
+        attn_dropout=0.1,
+        ff_dropout=0.1,
+        flash_attn=True,
+        dim_freqs_in=1025,
+        sample_rate=44100,  # needed for mel filter bank from librosa
+        stft_n_fft=2048,
+        stft_hop_length=512,
+        # 10ms at 44100Hz, from sections 4.1, 4.4 in the paper - @faroit recommends // 2 or // 4 for better reconstruction
+        stft_win_length=2048,
+        stft_normalized=False,
+        stft_window_fn: Optional[Callable] = None,
+        mask_estimator_depth=1,
+        multi_stft_resolution_loss_weight=1.0,
+        multi_stft_resolutions_window_sizes: Tuple[int, ...] = (4096, 2048, 1024, 512, 256),
+        multi_stft_hop_size=147,
+        multi_stft_normalized=False,
+        multi_stft_window_fn: Callable = torch.hann_window,
+        match_input_audio_length=False,  # if True, pad output tensor to match length of input tensor
+        mlp_expansion_factor=4,
+        use_torch_checkpoint=False,
+        skip_connection=False,
+    ):
+        super().__init__()
+
+        self.stereo = stereo
+        self.audio_channels = 2 if stereo else 1
+        self.num_stems = num_stems
+        self.use_torch_checkpoint = use_torch_checkpoint
+        self.skip_connection = skip_connection
+
+        self.layers = ModuleList([])
+
+        transformer_kwargs = dict(
+            dim=dim,
+            heads=heads,
+            dim_head=dim_head,
+            attn_dropout=attn_dropout,
+            ff_dropout=ff_dropout,
+            flash_attn=flash_attn,
+        )
+
+        time_rotary_embed = RotaryEmbedding(dim=dim_head)
+        freq_rotary_embed = RotaryEmbedding(dim=dim_head)
+
+        for _ in range(depth):
+            tran_modules = []
+            if linear_transformer_depth > 0:
+                tran_modules.append(Transformer(depth=linear_transformer_depth, linear_attn=True, **transformer_kwargs))
+            tran_modules.append(
+                Transformer(depth=time_transformer_depth, rotary_embed=time_rotary_embed, **transformer_kwargs)
+            )
+            tran_modules.append(
+                Transformer(depth=freq_transformer_depth, rotary_embed=freq_rotary_embed, **transformer_kwargs)
+            )
+            self.layers.append(nn.ModuleList(tran_modules))
+
+        self.stft_window_fn = partial(default(stft_window_fn, torch.hann_window), stft_win_length)
+
+        self.stft_kwargs = dict(
+            n_fft=stft_n_fft, hop_length=stft_hop_length, win_length=stft_win_length, normalized=stft_normalized
+        )
+
+        freqs = torch.stft(
+            torch.randn(1, 4096), **self.stft_kwargs, window=torch.ones(stft_n_fft), return_complex=True
+        ).shape[1]
+
+        # create mel filter bank
+        # with librosa.filters.mel as in section 2 of paper
+
+        mel_filter_bank_numpy = filters.mel(sr=sample_rate, n_fft=stft_n_fft, n_mels=num_bands)
+
+        mel_filter_bank = torch.from_numpy(mel_filter_bank_numpy)
+
+        # for some reason, it doesn't include the first freq? just force a value for now
+
+        mel_filter_bank[0][0] = 1.0
+
+        # In some systems/envs we get 0.0 instead of ~1.9e-18 in the last position,
+        # so let's force a positive value
+
+        mel_filter_bank[-1, -1] = 1.0
+
+        # binary as in paper (then estimated masks are averaged for overlapping regions)
+
+        freqs_per_band = mel_filter_bank > 0
+        assert freqs_per_band.any(dim=0).all(), "all frequencies need to be covered by all bands for now"
+
+        repeated_freq_indices = repeat(torch.arange(freqs), "f -> b f", b=num_bands)
+        freq_indices = repeated_freq_indices[freqs_per_band]
+
+        if stereo:
+            freq_indices = repeat(freq_indices, "f -> f s", s=2)
+            freq_indices = freq_indices * 2 + torch.arange(2)
+            freq_indices = rearrange(freq_indices, "f s -> (f s)")
+
+        self.register_buffer("freq_indices", freq_indices, persistent=False)
+        self.register_buffer("freqs_per_band", freqs_per_band, persistent=False)
+
+        num_freqs_per_band = reduce(freqs_per_band, "b f -> b", "sum")
+        num_bands_per_freq = reduce(freqs_per_band, "b f -> f", "sum")
+
+        self.register_buffer("num_freqs_per_band", num_freqs_per_band, persistent=False)
+        self.register_buffer("num_bands_per_freq", num_bands_per_freq, persistent=False)
+
+        # band split and mask estimator
+
+        freqs_per_bands_with_complex = tuple(2 * f * self.audio_channels for f in num_freqs_per_band.tolist())
+
+        self.band_split = BandSplit(dim=dim, dim_inputs=freqs_per_bands_with_complex)
+
+        self.mask_estimators = nn.ModuleList([])
+
+        for _ in range(num_stems):
+            mask_estimator = MaskEstimator(
+                dim=dim,
+                dim_inputs=freqs_per_bands_with_complex,
+                depth=mask_estimator_depth,
+                mlp_expansion_factor=mlp_expansion_factor,
+            )
+
+            self.mask_estimators.append(mask_estimator)
+
+        # for the multi-resolution stft loss
+
+        self.multi_stft_resolution_loss_weight = multi_stft_resolution_loss_weight
+        self.multi_stft_resolutions_window_sizes = multi_stft_resolutions_window_sizes
+        self.multi_stft_n_fft = stft_n_fft
+        self.multi_stft_window_fn = multi_stft_window_fn
+
+        self.multi_stft_kwargs = dict(hop_length=multi_stft_hop_size, normalized=multi_stft_normalized)
+
+        self.match_input_audio_length = match_input_audio_length
+
+    def forward(self, raw_audio, target=None, return_loss_breakdown=False):
+        """
+        einops
+
+        b - batch
+        f - freq
+        t - time
+        s - audio channel (1 for mono, 2 for stereo)
+        n - number of 'stems'
+        c - complex (2)
+        d - feature dimension
+        """
+
+        device = raw_audio.device
+
+        if raw_audio.ndim == 2:
+            raw_audio = rearrange(raw_audio, "b t -> b 1 t")
+
+        batch, channels, raw_audio_length = raw_audio.shape
+
+        istft_length = raw_audio_length if self.match_input_audio_length else None
+
+        assert (not self.stereo and channels == 1) or (self.stereo and channels == 2), (
+            "stereo needs to be set to True if passing in audio signal that is stereo (channel dimension of 2). also need to be False if mono (channel dimension of 1)"
+        )
+
+        # to stft
+
+        raw_audio, batch_audio_channel_packed_shape = pack_one(raw_audio, "* t")
+
+        stft_window = self.stft_window_fn(device=device)
+
+        stft_repr = torch.stft(raw_audio, **self.stft_kwargs, window=stft_window, return_complex=True)
+        stft_repr = torch.view_as_real(stft_repr)
+
+        stft_repr = unpack_one(stft_repr, batch_audio_channel_packed_shape, "* f t c")
+
+        # merge stereo / mono into the frequency, with frequency leading dimension, for band splitting
+        stft_repr = rearrange(stft_repr, "b s f t c -> b (f s) t c")
+
+        # index out all frequencies for all frequency ranges across bands ascending in one go
+
+        batch_arange = torch.arange(batch, device=device)[..., None]
+
+        # account for stereo
+
+        x = stft_repr[batch_arange, self.freq_indices]
+
+        # fold the complex (real and imag) into the frequencies dimension
+
+        x = rearrange(x, "b f t c -> b t (f c)")
+
+        if self.use_torch_checkpoint:
+            x = checkpoint(self.band_split, x, use_reentrant=False)
+        else:
+            x = self.band_split(x)
+
+        # axial / hierarchical attention
+
+        store = [None] * len(self.layers)
+        for i, transformer_block in enumerate(self.layers):
+            if len(transformer_block) == 3:
+                linear_transformer, time_transformer, freq_transformer = transformer_block
+
+                x, ft_ps = pack([x], "b * d")
+                if self.use_torch_checkpoint:
+                    x = checkpoint(linear_transformer, x, use_reentrant=False)
+                else:
+                    x = linear_transformer(x)
+                (x,) = unpack(x, ft_ps, "b * d")
+            else:
+                time_transformer, freq_transformer = transformer_block
+
+            if self.skip_connection:
+                # Sum all previous
+                for j in range(i):
+                    x = x + store[j]
+
+            x = rearrange(x, "b t f d -> b f t d")
+            x, ps = pack([x], "* t d")
+
+            if self.use_torch_checkpoint:
+                x = checkpoint(time_transformer, x, use_reentrant=False)
+            else:
+                x = time_transformer(x)
+
+            (x,) = unpack(x, ps, "* t d")
+            x = rearrange(x, "b f t d -> b t f d")
+            x, ps = pack([x], "* f d")
+
+            if self.use_torch_checkpoint:
+                x = checkpoint(freq_transformer, x, use_reentrant=False)
+            else:
+                x = freq_transformer(x)
+
+            (x,) = unpack(x, ps, "* f d")
+
+            if self.skip_connection:
+                store[i] = x
+
+        num_stems = len(self.mask_estimators)
+        if self.use_torch_checkpoint:
+            masks = torch.stack([checkpoint(fn, x, use_reentrant=False) for fn in self.mask_estimators], dim=1)
+        else:
+            masks = torch.stack([fn(x) for fn in self.mask_estimators], dim=1)
+        masks = rearrange(masks, "b n t (f c) -> b n f t c", c=2)
+
+        # modulate frequency representation
+
+        stft_repr = rearrange(stft_repr, "b f t c -> b 1 f t c")
+
+        # complex number multiplication
+
+        stft_repr = torch.view_as_complex(stft_repr)
+        masks = torch.view_as_complex(masks)
+
+        masks = masks.type(stft_repr.dtype)
+
+        # need to average the estimated mask for the overlapped frequencies
+
+        scatter_indices = repeat(self.freq_indices, "f -> b n f t", b=batch, n=num_stems, t=stft_repr.shape[-1])
+
+        stft_repr_expanded_stems = repeat(stft_repr, "b 1 ... -> b n ...", n=num_stems)
+        masks_summed = torch.zeros_like(stft_repr_expanded_stems).scatter_add_(2, scatter_indices, masks)
+
+        denom = repeat(self.num_bands_per_freq, "f -> (f r) 1", r=channels)
+
+        masks_averaged = masks_summed / denom.clamp(min=1e-8)
+
+        # modulate stft repr with estimated mask
+
+        stft_repr = stft_repr * masks_averaged
+
+        # istft
+
+        stft_repr = rearrange(stft_repr, "b n (f s) t -> (b n s) f t", s=self.audio_channels)
+
+        recon_audio = torch.istft(
+            stft_repr, **self.stft_kwargs, window=stft_window, return_complex=False, length=istft_length
+        )
+
+        recon_audio = rearrange(recon_audio, "(b n s) t -> b n s t", b=batch, s=self.audio_channels, n=num_stems)
+
+        if num_stems == 1:
+            recon_audio = rearrange(recon_audio, "b 1 s t -> b s t")
+
+        # if a target is passed in, calculate loss for learning
+
+        if not exists(target):
+            return recon_audio
+
+        if self.num_stems > 1:
+            assert target.ndim == 4 and target.shape[1] == self.num_stems
+
+        if target.ndim == 2:
+            target = rearrange(target, "... t -> ... 1 t")
+
+        target = target[..., : recon_audio.shape[-1]]  # protect against lost length on istft
+
+        loss = F.l1_loss(recon_audio, target)
+
+        multi_stft_resolution_loss = 0.0
+
+        for window_size in self.multi_stft_resolutions_window_sizes:
+            res_stft_kwargs = dict(
+                n_fft=max(window_size, self.multi_stft_n_fft),  # not sure what n_fft is across multi resolution stft
+                win_length=window_size,
+                return_complex=True,
+                window=self.multi_stft_window_fn(window_size, device=device),
+                **self.multi_stft_kwargs,
+            )
+
+            recon_Y = torch.stft(rearrange(recon_audio, "... s t -> (... s) t"), **res_stft_kwargs)
+            target_Y = torch.stft(rearrange(target, "... s t -> (... s) t"), **res_stft_kwargs)
+
+            multi_stft_resolution_loss = multi_stft_resolution_loss + F.l1_loss(recon_Y, target_Y)
+
+        weighted_multi_resolution_loss = multi_stft_resolution_loss * self.multi_stft_resolution_loss_weight
+
+        total_loss = loss + weighted_multi_resolution_loss
+
+        if not return_loss_breakdown:
+            return total_loss
+
+        return total_loss, (loss, multi_stft_resolution_loss)
diff --git a/tools/uvr5/bsroformer.py b/tools/uvr5/bsroformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..ddcbfa749e663e1f546db448b81cf38be4018860
--- /dev/null
+++ b/tools/uvr5/bsroformer.py
@@ -0,0 +1,304 @@
+# This code is modified from https://github.com/ZFTurbo/
+import os
+import warnings
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+import torch.nn as nn
+import yaml
+from tqdm import tqdm
+
+warnings.filterwarnings("ignore")
+
+
+class Roformer_Loader:
+    def get_config(self, config_path):
+        with open(config_path, "r", encoding="utf-8") as f:
+            # use fullloader to load tag !!python/tuple, code can be improved
+            config = yaml.load(f, Loader=yaml.FullLoader)
+        return config
+
+    def get_default_config(self):
+        default_config = None
+        if self.model_type == "bs_roformer":
+            # Use model_bs_roformer_ep_368_sdr_12.9628.yaml and model_bs_roformer_ep_317_sdr_12.9755.yaml as default configuration files
+            # Other BS_Roformer models may not be compatible
+            # fmt: off
+            default_config = {
+                "audio": {"chunk_size": 352800, "sample_rate": 44100},
+                "model": {
+                    "dim": 512,
+                    "depth": 12,
+                    "stereo": True,
+                    "num_stems": 1,
+                    "time_transformer_depth": 1,
+                    "freq_transformer_depth": 1,
+                    "linear_transformer_depth": 0,
+                    "freqs_per_bands": (2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 12, 12, 12, 12, 12, 12, 12, 12, 24, 24, 24, 24, 24, 24, 24, 24, 48, 48, 48, 48, 48, 48, 48, 48, 128, 129),
+                    "dim_head": 64,
+                    "heads": 8,
+                    "attn_dropout": 0.1,
+                    "ff_dropout": 0.1,
+                    "flash_attn": True,
+                    "dim_freqs_in": 1025,
+                    "stft_n_fft": 2048,
+                    "stft_hop_length": 441,
+                    "stft_win_length": 2048,
+                    "stft_normalized": False,
+                    "mask_estimator_depth": 2,
+                    "multi_stft_resolution_loss_weight": 1.0,
+                    "multi_stft_resolutions_window_sizes": (4096, 2048, 1024, 512, 256),
+                    "multi_stft_hop_size": 147,
+                    "multi_stft_normalized": False,
+                },
+                "training": {"instruments": ["vocals", "other"], "target_instrument": "vocals"},
+                "inference": {"batch_size": 2, "num_overlap": 2},
+            }
+            # fmt: on
+        elif self.model_type == "mel_band_roformer":
+            # Use model_mel_band_roformer_ep_3005_sdr_11.4360.yaml as default configuration files
+            # Other Mel_Band_Roformer models may not be compatible
+            default_config = {
+                "audio": {"chunk_size": 352800, "sample_rate": 44100},
+                "model": {
+                    "dim": 384,
+                    "depth": 12,
+                    "stereo": True,
+                    "num_stems": 1,
+                    "time_transformer_depth": 1,
+                    "freq_transformer_depth": 1,
+                    "linear_transformer_depth": 0,
+                    "num_bands": 60,
+                    "dim_head": 64,
+                    "heads": 8,
+                    "attn_dropout": 0.1,
+                    "ff_dropout": 0.1,
+                    "flash_attn": True,
+                    "dim_freqs_in": 1025,
+                    "sample_rate": 44100,
+                    "stft_n_fft": 2048,
+                    "stft_hop_length": 441,
+                    "stft_win_length": 2048,
+                    "stft_normalized": False,
+                    "mask_estimator_depth": 2,
+                    "multi_stft_resolution_loss_weight": 1.0,
+                    "multi_stft_resolutions_window_sizes": (4096, 2048, 1024, 512, 256),
+                    "multi_stft_hop_size": 147,
+                    "multi_stft_normalized": False,
+                },
+                "training": {"instruments": ["vocals", "other"], "target_instrument": "vocals"},
+                "inference": {"batch_size": 2, "num_overlap": 2},
+            }
+
+        return default_config
+
+    def get_model_from_config(self):
+        if self.model_type == "bs_roformer":
+            from bs_roformer.bs_roformer import BSRoformer
+
+            model = BSRoformer(**dict(self.config["model"]))
+        elif self.model_type == "mel_band_roformer":
+            from bs_roformer.mel_band_roformer import MelBandRoformer
+
+            model = MelBandRoformer(**dict(self.config["model"]))
+        else:
+            print("Error: Unknown model: {}".format(self.model_type))
+            model = None
+        return model
+
+    def demix_track(self, model, mix, device):
+        C = self.config["audio"]["chunk_size"]  # chunk_size
+        N = self.config["inference"]["num_overlap"]
+        fade_size = C // 10
+        step = int(C // N)
+        border = C - step
+        batch_size = self.config["inference"]["batch_size"]
+
+        length_init = mix.shape[-1]
+        progress_bar = tqdm(total=length_init // step + 1, desc="Processing", leave=False)
+
+        # Do pad from the beginning and end to account floating window results better
+        if length_init > 2 * border and (border > 0):
+            mix = nn.functional.pad(mix, (border, border), mode="reflect")
+
+        # Prepare windows arrays (do 1 time for speed up). This trick repairs click problems on the edges of segment
+        window_size = C
+        fadein = torch.linspace(0, 1, fade_size)
+        fadeout = torch.linspace(1, 0, fade_size)
+        window_start = torch.ones(window_size)
+        window_middle = torch.ones(window_size)
+        window_finish = torch.ones(window_size)
+        window_start[-fade_size:] *= fadeout  # First audio chunk, no fadein
+        window_finish[:fade_size] *= fadein  # Last audio chunk, no fadeout
+        window_middle[-fade_size:] *= fadeout
+        window_middle[:fade_size] *= fadein
+
+        with torch.amp.autocast("cuda"):
+            with torch.inference_mode():
+                if self.config["training"]["target_instrument"] is None:
+                    req_shape = (len(self.config["training"]["instruments"]),) + tuple(mix.shape)
+                else:
+                    req_shape = (1,) + tuple(mix.shape)
+
+                result = torch.zeros(req_shape, dtype=torch.float32)
+                counter = torch.zeros(req_shape, dtype=torch.float32)
+                i = 0
+                batch_data = []
+                batch_locations = []
+                while i < mix.shape[1]:
+                    part = mix[:, i : i + C].to(device)
+                    length = part.shape[-1]
+                    if length < C:
+                        if length > C // 2 + 1:
+                            part = nn.functional.pad(input=part, pad=(0, C - length), mode="reflect")
+                        else:
+                            part = nn.functional.pad(input=part, pad=(0, C - length, 0, 0), mode="constant", value=0)
+                    if self.is_half:
+                        part = part.half()
+                    batch_data.append(part)
+                    batch_locations.append((i, length))
+                    i += step
+                    progress_bar.update(1)
+
+                    if len(batch_data) >= batch_size or (i >= mix.shape[1]):
+                        arr = torch.stack(batch_data, dim=0)
+                        # print(23333333,arr.dtype)
+                        x = model(arr)
+
+                        window = window_middle
+                        if i - step == 0:  # First audio chunk, no fadein
+                            window = window_start
+                        elif i >= mix.shape[1]:  # Last audio chunk, no fadeout
+                            window = window_finish
+
+                        for j in range(len(batch_locations)):
+                            start, l = batch_locations[j]
+                            result[..., start : start + l] += x[j][..., :l].cpu() * window[..., :l]
+                            counter[..., start : start + l] += window[..., :l]
+
+                        batch_data = []
+                        batch_locations = []
+
+                estimated_sources = result / counter
+                estimated_sources = estimated_sources.cpu().numpy()
+                np.nan_to_num(estimated_sources, copy=False, nan=0.0)
+
+                if length_init > 2 * border and (border > 0):
+                    # Remove pad
+                    estimated_sources = estimated_sources[..., border:-border]
+
+        progress_bar.close()
+
+        if self.config["training"]["target_instrument"] is None:
+            return {k: v for k, v in zip(self.config["training"]["instruments"], estimated_sources)}
+        else:
+            return {k: v for k, v in zip([self.config["training"]["target_instrument"]], estimated_sources)}
+
+    def run_folder(self, input, vocal_root, others_root, format):
+        self.model.eval()
+        path = input
+        os.makedirs(vocal_root, exist_ok=True)
+        os.makedirs(others_root, exist_ok=True)
+        file_base_name = os.path.splitext(os.path.basename(path))[0]
+
+        sample_rate = 44100
+        if "sample_rate" in self.config["audio"]:
+            sample_rate = self.config["audio"]["sample_rate"]
+
+        try:
+            mix, sr = librosa.load(path, sr=sample_rate, mono=False)
+        except Exception as e:
+            print("Can read track: {}".format(path))
+            print("Error message: {}".format(str(e)))
+            return
+
+        # in case if model only supports mono tracks
+        isstereo = self.config["model"].get("stereo", True)
+        if not isstereo and len(mix.shape) != 1:
+            mix = np.mean(mix, axis=0)  # if more than 2 channels, take mean
+            print("Warning: Track has more than 1 channels, but model is mono, taking mean of all channels.")
+
+        mix_orig = mix.copy()
+
+        mixture = torch.tensor(mix, dtype=torch.float32)
+        res = self.demix_track(self.model, mixture, self.device)
+
+        if self.config["training"]["target_instrument"] is not None:
+            # if target instrument is specified, save target instrument as vocal and other instruments as others
+            # other instruments are caculated by subtracting target instrument from mixture
+            target_instrument = self.config["training"]["target_instrument"]
+            other_instruments = [i for i in self.config["training"]["instruments"] if i != target_instrument]
+            other = mix_orig - res[target_instrument]  # caculate other instruments
+
+            path_vocal = "{}/{}_{}.wav".format(vocal_root, file_base_name, target_instrument)
+            path_other = "{}/{}_{}.wav".format(others_root, file_base_name, other_instruments[0])
+            self.save_audio(path_vocal, res[target_instrument].T, sr, format)
+            self.save_audio(path_other, other.T, sr, format)
+        else:
+            # if target instrument is not specified, save the first instrument as vocal and the rest as others
+            vocal_inst = self.config["training"]["instruments"][0]
+            path_vocal = "{}/{}_{}.wav".format(vocal_root, file_base_name, vocal_inst)
+            self.save_audio(path_vocal, res[vocal_inst].T, sr, format)
+            for other in self.config["training"]["instruments"][1:]:  # save other instruments
+                path_other = "{}/{}_{}.wav".format(others_root, file_base_name, other)
+                self.save_audio(path_other, res[other].T, sr, format)
+
+    def save_audio(self, path, data, sr, format):
+        # input path should be endwith '.wav'
+        if format in ["wav", "flac"]:
+            if format == "flac":
+                path = path[:-3] + "flac"
+            sf.write(path, data, sr)
+        else:
+            sf.write(path, data, sr)
+            os.system('ffmpeg -i "{}" -vn "{}" -q:a 2 -y'.format(path, path[:-3] + format))
+            try:
+                os.remove(path)
+            except:
+                pass
+
+    def __init__(self, model_path, config_path, device, is_half):
+        self.device = device
+        self.is_half = is_half
+        self.model_type = None
+        self.config = None
+
+        # get model_type, first try:
+        if "bs_roformer" in model_path.lower() or "bsroformer" in model_path.lower():
+            self.model_type = "bs_roformer"
+        elif "mel_band_roformer" in model_path.lower() or "melbandroformer" in model_path.lower():
+            self.model_type = "mel_band_roformer"
+
+        if not os.path.exists(config_path):
+            if self.model_type is None:
+                # if model_type is still None, raise an error
+                raise ValueError(
+                    "Error: Unknown model type. If you are using a model without a configuration file, Ensure that your model name includes 'bs_roformer', 'bsroformer', 'mel_band_roformer', or 'melbandroformer'. Otherwise, you can manually place the model configuration file into 'tools/uvr5/uvr5w_weights' and ensure that the configuration file is named as '<model_name>.yaml' then try it again."
+                )
+            self.config = self.get_default_config()
+        else:
+            # if there is a configuration file
+            self.config = self.get_config(config_path)
+            if self.model_type is None:
+                # if model_type is still None, second try, get model_type from the configuration file
+                if "freqs_per_bands" in self.config["model"]:
+                    # if freqs_per_bands in config, it's a bs_roformer model
+                    self.model_type = "bs_roformer"
+                else:
+                    # else it's a mel_band_roformer model
+                    self.model_type = "mel_band_roformer"
+
+        print("Detected model type: {}".format(self.model_type))
+        model = self.get_model_from_config()
+        state_dict = torch.load(model_path, map_location="cpu")
+        model.load_state_dict(state_dict)
+
+        if is_half == False:
+            self.model = model.to(device)
+        else:
+            self.model = model.half().to(device)
+
+    def _path_audio_(self, input, others_root, vocal_root, format, is_hp3=False):
+        self.run_folder(input, vocal_root, others_root, format)
diff --git a/tools/uvr5/mdxnet.py b/tools/uvr5/mdxnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..e109827343cbcd6c3375e12bc02b15c5f38d2d11
--- /dev/null
+++ b/tools/uvr5/mdxnet.py
@@ -0,0 +1,223 @@
+import os
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+from tqdm import tqdm
+
+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
+            if os.path.exists(path_vocal):
+                os.system("ffmpeg -i '%s' -vn '%s' -q:a 2 -y" % (path_vocal, opt_path_vocal))
+                if os.path.exists(opt_path_vocal):
+                    try:
+                        os.remove(path_vocal)
+                    except:
+                        pass
+            if os.path.exists(path_other):
+                os.system("ffmpeg -i '%s' -vn '%s' -q:a 2 -y" % (path_other, opt_path_other))
+                if os.path.exists(opt_path_other):
+                    try:
+                        os.remove(path_other)
+                    except:
+                        pass
+
+
+class MDXNetDereverb:
+    def __init__(self, chunks):
+        self.onnx = "%s/uvr5_weights/onnx_dereverb_By_FoxJoy" % os.path.dirname(os.path.abspath(__file__))
+        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 = cpu
+
+    def _path_audio_(self, input, others_root, vocal_root, format, is_hp3=False):
+        self.pred.prediction(input, vocal_root, others_root, format)
diff --git a/tools/uvr5/uvr5_weights/.gitignore b/tools/uvr5/uvr5_weights/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..d6b7ef32c8478a48c3994dcadc86837f4371184d
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/.gitignore
@@ -0,0 +1,2 @@
+*
+!.gitignore
diff --git a/tools/uvr5/uvr5_weights/HP2_all_vocals.pth b/tools/uvr5/uvr5_weights/HP2_all_vocals.pth
new file mode 100644
index 0000000000000000000000000000000000000000..2594ca1392e5e9f7b8d9a8535c7174bf36309a9e
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/HP2_all_vocals.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39796caa5db18d7f9382d8ac997ac967bfd85f7761014bb807d2543cc844ef05
+size 63454827
diff --git a/tools/uvr5/uvr5_weights/HP5_only_main_vocal.pth b/tools/uvr5/uvr5_weights/HP5_only_main_vocal.pth
new file mode 100644
index 0000000000000000000000000000000000000000..37265088b4cff876d7635addd70ffa3f0fb259e1
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/HP5_only_main_vocal.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5908891829634926119720241e8573d97cbeb8277110a7512bdb0bd7563258ee
+size 63454827
diff --git a/tools/uvr5/uvr5_weights/VR-DeEchoAggressive.pth b/tools/uvr5/uvr5_weights/VR-DeEchoAggressive.pth
new file mode 100644
index 0000000000000000000000000000000000000000..972d24d51d322c31643e846c81c5f05096426324
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/VR-DeEchoAggressive.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c8fd1582f9aabc363e47af62ddb88df6cae7e064cae75bbf041a067a5e0aee2
+size 63666335
diff --git a/tools/uvr5/uvr5_weights/VR-DeEchoDeReverb.pth b/tools/uvr5/uvr5_weights/VR-DeEchoDeReverb.pth
new file mode 100644
index 0000000000000000000000000000000000000000..7c7820029b25f35bc05a0ff8dbc0936a39dc0577
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/VR-DeEchoDeReverb.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:01376dd2a571bf3cb9cced680732726d2d732609d09216a610b0d110f133febe
+size 111925279
diff --git a/tools/uvr5/uvr5_weights/VR-DeEchoNormal.pth b/tools/uvr5/uvr5_weights/VR-DeEchoNormal.pth
new file mode 100644
index 0000000000000000000000000000000000000000..7b702f38fc1009306b806e8a8fdbbf5d2a59fb4d
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/VR-DeEchoNormal.pth
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56aba59db3bcdd14a14464e62f3129698ecdea62eee0f003b9360923eb3ac79e
+size 63666335
diff --git a/tools/uvr5/uvr5_weights/model_bs_roformer_ep_317_sdr_12.9755.ckpt b/tools/uvr5/uvr5_weights/model_bs_roformer_ep_317_sdr_12.9755.ckpt
new file mode 100644
index 0000000000000000000000000000000000000000..6a28253d1d71bd88c9133c45b9b20d2d9fa59c62
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/model_bs_roformer_ep_317_sdr_12.9755.ckpt
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:414b7ea8e3f57485282886f0bae273a898e02b9ca14524cdefa0fb9f5784e4d1
+size 319754015
diff --git a/tools/uvr5/uvr5_weights/onnx_dereverb_By_FoxJoy/vocals.onnx b/tools/uvr5/uvr5_weights/onnx_dereverb_By_FoxJoy/vocals.onnx
new file mode 100644
index 0000000000000000000000000000000000000000..0af4086572259731e9237a7d5acc2254f2d30cf5
--- /dev/null
+++ b/tools/uvr5/uvr5_weights/onnx_dereverb_By_FoxJoy/vocals.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:233bb5c6aaa365e568659a0a81211746fa881f8f47f82d9e864fce1f7692db80
+size 66780123
diff --git a/tools/uvr5/vr.py b/tools/uvr5/vr.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ca8a3b780c4870861166a475145e4b120c65d3d
--- /dev/null
+++ b/tools/uvr5/vr.py
@@ -0,0 +1,342 @@
+import os
+
+parent_directory = os.path.dirname(os.path.abspath(__file__))
+import logging
+
+logger = logging.getLogger(__name__)
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+from lib.lib_v5 import nets_61968KB as Nets
+from lib.lib_v5 import spec_utils
+from lib.lib_v5.model_param_init import ModelParameters
+from lib.lib_v5.nets_new import CascadedNet
+from lib.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("%s/lib/lib_v5/modelparams/4band_v2.json" % parent_directory)
+        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ffmpeg读取，但是太麻烦了弃坑
+                    music_file,
+                    sr=bp["sr"],
+                    mono=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],
+                    orig_sr=self.mp.param["band"][d + 1]["sr"],
+                    target_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 is_hp3 == True:
+            ins_root, vocal_root = vocal_root, ins_root
+
+        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
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
+        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"],
+                )
+                if os.path.exists(path):
+                    opt_format_path = path[:-4] + ".%s" % format
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
+
+
+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("%s/lib/lib_v5/modelparams/4band_v3.json" % parent_directory)
+        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ffmpeg读取，但是太麻烦了弃坑
+                    music_file,
+                    sr=bp["sr"],
+                    mono=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],
+                    orig_sr=self.mp.param["band"][d + 1]["sr"],
+                    target_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
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
+        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
+                    os.system("ffmpeg -i %s -vn %s -q:a 2 -y" % (path, opt_format_path))
+                    if os.path.exists(opt_format_path):
+                        try:
+                            os.remove(path)
+                        except:
+                            pass
diff --git a/tools/uvr5/webui.py b/tools/uvr5/webui.py
new file mode 100644
index 0000000000000000000000000000000000000000..cc826bf238ee67d44eddf6e24d39e2a0d9c74802
--- /dev/null
+++ b/tools/uvr5/webui.py
@@ -0,0 +1,228 @@
+import os
+import traceback
+import gradio as gr
+import logging
+from tools.i18n.i18n import I18nAuto
+from tools.my_utils import clean_path
+
+i18n = I18nAuto()
+
+logger = logging.getLogger(__name__)
+import ffmpeg
+import torch
+import sys
+from mdxnet import MDXNetDereverb
+from vr import AudioPre, AudioPreDeEcho
+from bsroformer import Roformer_Loader
+
+try:
+    import gradio.analytics as analytics
+
+    analytics.version_check = lambda: None
+except:
+    ...
+
+weight_uvr5_root = "tools/uvr5/uvr5_weights"
+uvr5_names = []
+for name in os.listdir(weight_uvr5_root):
+    if name.endswith(".pth") or name.endswith(".ckpt") or "onnx" in name:
+        uvr5_names.append(name.replace(".pth", "").replace(".ckpt", ""))
+
+device = sys.argv[1]
+is_half = eval(sys.argv[2])
+webui_port_uvr5 = int(sys.argv[3])
+is_share = eval(sys.argv[4])
+
+
+def html_left(text, label="p"):
+    return f"""<div style="text-align: left; margin: 0; padding: 0;">
+                <{label} style="margin: 0; padding: 0;">{text}</{label}>
+                </div>"""
+
+
+def html_center(text, label="p"):
+    return f"""<div style="text-align: center; margin: 100; padding: 50;">
+                <{label} style="margin: 0; padding: 0;">{text}</{label}>
+                </div>"""
+
+
+def uvr(model_name, inp_root, save_root_vocal, paths, save_root_ins, agg, format0):
+    infos = []
+    try:
+        inp_root = clean_path(inp_root)
+        save_root_vocal = clean_path(save_root_vocal)
+        save_root_ins = clean_path(save_root_ins)
+        is_hp3 = "HP3" in model_name
+        if model_name == "onnx_dereverb_By_FoxJoy":
+            pre_fun = MDXNetDereverb(15)
+        elif "roformer" in model_name.lower():
+            func = Roformer_Loader
+            pre_fun = func(
+                model_path=os.path.join(weight_uvr5_root, model_name + ".ckpt"),
+                config_path=os.path.join(weight_uvr5_root, model_name + ".yaml"),
+                device=device,
+                is_half=is_half,
+            )
+            if not os.path.exists(os.path.join(weight_uvr5_root, model_name + ".yaml")):
+                infos.append(
+                    "Warning: You are using a model without a configuration file. The program will automatically use the default configuration file. However, the default configuration file cannot guarantee that all models will run successfully. You can manually place the model configuration file into 'tools/uvr5/uvr5w_weights' and ensure that the configuration file is named as '<model_name>.yaml' then try it again. (For example, the configuration file corresponding to the model 'bs_roformer_ep_368_sdr_12.9628.ckpt' should be 'bs_roformer_ep_368_sdr_12.9628.yaml'.) Or you can just ignore this warning."
+                )
+                yield "\n".join(infos)
+        else:
+            func = AudioPre if "DeEcho" not in model_name else AudioPreDeEcho
+            pre_fun = func(
+                agg=int(agg),
+                model_path=os.path.join(weight_uvr5_root, model_name + ".pth"),
+                device=device,
+                is_half=is_half,
+            )
+        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)
+            if os.path.isfile(inp_path) == False:
+                continue
+            need_reformat = 1
+            done = 0
+            try:
+                info = ffmpeg.probe(inp_path, cmd="ffprobe")
+                if info["streams"][0]["channels"] == 2 and info["streams"][0]["sample_rate"] == "44100":
+                    need_reformat = 0
+                    pre_fun._path_audio_(inp_path, save_root_ins, save_root_vocal, format0, is_hp3)
+                    done = 1
+            except:
+                need_reformat = 1
+                traceback.print_exc()
+            if need_reformat == 1:
+                tmp_path = "%s/%s.reformatted.wav" % (
+                    os.path.join(os.environ["TEMP"]),
+                    os.path.basename(inp_path),
+                )
+                os.system(f'ffmpeg -i "{inp_path}" -vn -acodec pcm_s16le -ac 2 -ar 44100 "{tmp_path}" -y')
+                inp_path = tmp_path
+            try:
+                if done == 0:
+                    pre_fun._path_audio_(inp_path, save_root_ins, save_root_vocal, format0, is_hp3)
+                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()
+        print("clean_empty_cache")
+        if torch.cuda.is_available():
+            torch.cuda.empty_cache()
+    yield "\n".join(infos)
+
+
+with gr.Blocks(title="UVR5 WebUI") as app:
+    gr.Markdown(
+        value=i18n("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.")
+        + "<br>"
+        + i18n("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.")
+    )
+    with gr.Group():
+        gr.Markdown(html_center(i18n("伴奏人声分离&去混响&去回声"), "h2"))
+        with gr.Group():
+            gr.Markdown(
+                value=html_left(
+                    i18n("人声伴奏分离批量处理， 使用UVR5模型。")
+                    + "<br>"
+                    + i18n(
+                        "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)。"
+                    )
+                    + "<br>"
+                    + i18n("模型分为三类：")
+                    + "<br>"
+                    + i18n(
+                        "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；"
+                    )
+                    + "<br>"
+                    + i18n("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；")
+                    + "<br>"
+                    + i18n("3、去混响、去延迟模型（by FoxJoy）：")
+                    + "<br>  "
+                    + i18n("(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；")
+                    + "<br>&emsp;"
+                    + i18n(
+                        "(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。"
+                    )
+                    + "<br>"
+                    + i18n("去混响/去延迟，附：")
+                    + "<br>"
+                    + i18n("1、DeEcho-DeReverb模型的耗时是另外2个DeEcho模型的接近2倍；")
+                    + "<br>"
+                    + i18n("2、MDX-Net-Dereverb模型挺慢的；")
+                    + "<br>"
+                    + i18n("3、个人推荐的最干净的配置是先MDX-Net再DeEcho-Aggressive。"),
+                    "h4",
+                )
+            )
+            with gr.Row():
+                with gr.Column():
+                    model_choose = gr.Dropdown(label=i18n("模型"), choices=uvr5_names)
+                    dir_wav_input = gr.Textbox(
+                        label=i18n("输入待处理音频文件夹路径"),
+                        placeholder="C:\\Users\\Desktop\\todo-songs",
+                    )
+                    wav_inputs = gr.File(
+                        file_count="multiple", label=i18n("也可批量输入音频文件, 二选一, 优先读文件夹")
+                    )
+                with gr.Column():
+                    agg = gr.Slider(
+                        minimum=0,
+                        maximum=20,
+                        step=1,
+                        label=i18n("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("指定输出主人声文件夹"), value="output/uvr5_opt")
+                    opt_ins_root = gr.Textbox(label=i18n("指定输出非主人声文件夹"), value="output/uvr5_opt")
+                    format0 = gr.Radio(
+                        label=i18n("导出文件格式"),
+                        choices=["wav", "flac", "mp3", "m4a"],
+                        value="flac",
+                        interactive=True,
+                    )
+                    with gr.Column():
+                        with gr.Row():
+                            but2 = gr.Button(i18n("转换"), variant="primary")
+                        with gr.Row():
+                            vc_output4 = gr.Textbox(label=i18n("输出信息"), lines=3)
+                but2.click(
+                    uvr,
+                    [
+                        model_choose,
+                        dir_wav_input,
+                        opt_vocal_root,
+                        wav_inputs,
+                        opt_ins_root,
+                        agg,
+                        format0,
+                    ],
+                    [vc_output4],
+                    api_name="uvr_convert",
+                )
+app.queue().launch(  # concurrency_count=511, max_size=1022
+    server_name="0.0.0.0",
+    inbrowser=True,
+    share=is_share,
+    server_port=webui_port_uvr5,
+    quiet=True,
+)