File size: 7,744 Bytes
ec04423 7e9e20b ec04423 871a2db c7f79fa 256f544 c7f79fa 871a2db c7f79fa 871a2db ec04423 256f544 ec04423 256f544 b2bfda4 e9d1415 ec04423 256f544 d0c88b1 ec04423 a9f3251 c2caf64 ec04423 a9f3251 c2caf64 a9f3251 0636449 7c25161 ec04423 29786e4 ec04423 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 |
---
license: other
license_name: nsclv1
license_link: https://developer.nvidia.com/downloads/license/nsclv1
---
# NVIDIA Low Frame-rate Speech Codec
<style>
img {
display: inline-table;
vertical-align: small;
margin: 0;
padding: 0;
}
</style>
[](#model-architecture)
| [](#model-architecture)
| [](#datasets)
The [Low Frame-rate Speech Codec](https://arxiv.org/abs/2409.12117) is a neural audio codec that leverages finite scalar quantization and adversarial training with large speech language models to achieve high-quality audio compression with a 1.89 kbps bitrate and 21.5 frames per second.
## Model Architecture
Low Frame-rate Speech Codec model is composed of a fully convolutional generator neural network and three discriminators.
The generator comprises an encoder, followed by vector quantization, and a [HiFi-GAN-based](https://arxiv.org/abs/2010.05646) decoder.
The encoder consists of five residual blocks, each block containing three residual layers similar to the [multi-receptive field fusion (MRF) module](https://arxiv.org/abs/2010.05646).
For the vector quantization, we have used [Finite Scalar Quantization (FSQ)](https://arxiv.org/abs/2309.15505) with eight codebooks and four dimensions per code and 2016 codes per codebook.
For the discriminators, we utilize three neural networks, all employing a squared-GAN and feature-matching loss. We adopt the [multi-period discriminator](https://arxiv.org/abs/2010.05646) and the [multi-scale complex STFT discriminator](https://arxiv.org/abs/2210.13438).
Additionally, we proposed the use of Speech Language Models (SLMs) as a discriminator. SLMs encode information ranging from acoustic to semantic aspects, which could benefit our model's training, especially in low frame rate settings where accurate pronunciation is difficult to achieve due to the high compression rate. We adopted the [12-layer WavLM](https://arxiv.org/abs/2110.13900), pre-trained on 94k hours of data, as the SLM. During training, we resample the input audio to 16 kHz before feeding it into the WavLM model, extracting the intermediary layer features. These features are then fed to a discriminative head composed of four 1D convolutional layers.
For more details please check [our paper](https://arxiv.org/abs/2409.12117).
### Input
- **Input Type:** Audio
- **Input Format(s):** .wav files
- **Input Parameters:** One-Dimensional (1D)
- **Other Properties Related to Input:** 22050 Hz Mono-channel Audio
### Output
- **Output Type**: Audio
- **Output Format:** .wav files
- **Output Parameters:** One Dimensional (1D)
- **Other Properties Related to Output:** 22050 Hz Mono-channel Audio
## How to Use this Model
The model is available for use in the [NVIDIA NeMo](https://github.com/NVIDIA/NeMo), and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
### Inference
For inference, you can manually download the [checkpoint](https://huggingface.co/nvidia/Low-Frame-rate-Speech-Codec/resolve/main/Low_Frame-rate_Speech_Codec_21Hz.nemo) and use the code below to make an inference on the model:
```
import librosa
import torch
import soundfile as sf
from nemo.collections.tts.models import AudioCodecModel
codec_path = ??? # set here the model .nemo checkpoint path
path_to_input_audio = ??? # path of the input audio
path_to_output_audio = ??? # path of the reconstructed output audio
nemo_codec_model = AudioCodecModel.restore_from(restore_path=codec_path, map_location="cpu").eval()
# get discrete tokens from audio
audio, _ = librosa.load(path_to_input_audio, sr=nemo_codec_model.sample_rate)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
audio_tensor = torch.from_numpy(audio).unsqueeze(dim=0).to(device)
audio_len = torch.tensor([audio_tensor[0].shape[0]]).to(device)
encoded_tokens, encoded_len = nemo_codec_model.encode(audio=audio_tensor, audio_len=audio_len)
# Reconstruct audio from tokens
reconstructed_audio, _ = nemo_codec_model.decode(tokens=encoded_tokens, tokens_len=encoded_len)
# save reconstructed audio
output_audio = reconstructed_audio.cpu().numpy().squeeze()
sf.write(path_to_output_audio, output_audio, nemo_codec_model.sample_rate)
```
### Training
For fine-tuning on another dataset please follow the steps available at our [Audio Codec Training Tutorial](https://github.com/NVIDIA/NeMo/blob/main/tutorials/tts/Audio_Codec_Training.ipynb). Note that you will need to set the ```CONFIG_FILENAME``` parameter to the "audio_codec_low_frame_rate_22050.yaml" config. You also will need to set ```pretrained_model_name``` to "audio_codec_low_frame_rate_22khz".
## Training, Testing, and Evaluation Datasets:
The Low Frame-rate Speech Codec was trained on 28.7k hours of speech data spanning 105 languages. The model was evaluated using multilingual audiobook-style data and high-quality English recordings. For further details, refer to [our paper](https://arxiv.org/abs/2409.12117).
### Training Datasets
The Low Frame-rate Speech Codec is trained on a total of 28.7k hrs of speech data from 105 languages.
- [MLS English](https://www.openslr.org/94/) [25.5k]
- Data Collection Method: by Human
- Labeling Method: Automated
- [Common Voice](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0)[3.2k]
- Data Collection Method: by Human
- Labeling Method: by Human
### Evaluation Datasets
- [MLS English](https://www.openslr.org/94/)
- Data Collection Method: by Human
- Labeling Method: Automated
- [Common Voice](https://huggingface.co/datasets/mozilla-foundation/common_voice_11_0)
- Data Collection Method: by Human
- Labeling Method: by Human
### Test Datasets
- [MLS](https://www.openslr.org/94/)
- Data Collection Method: by Human
- Labeling Method: Automated
- Properties: We randomly selected 200 samples from each of the eight languages in the 44kHz MLS dataset.
- [DAPS](https://zenodo.org/records/4660670)
- Data Collection Method: by Human
- Labeling Method: Automated
- Properties: To assess our models' performance on studio-quality audio, we utilized the F10 and M10 speakers from the DAPS Clear dataset. These speakers were also employed in the evaluation of the [DAC model](https://arxiv.org/abs/2306.06546).
## Software Integration
### Supported Hardware Microarchitecture Compatibility:
- NVIDIA Ampere
- NVIDIA Blackwell
- NVIDIA Jetson
- NVIDIA Hopper
- NVIDIA Lovelace
- NVIDIA Pascal
- NVIDIA Turing
- NVIDIA Volta
### Runtime Engine
- Nemo 2.0.0
### Preferred Operating System
- Linux
## License/Terms of Use
This model is for research and development only (non-commercial use) and the license to use this model is covered by the [NSCLv1](https://developer.nvidia.com/downloads/license/nsclv1).
## Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
|