Vocos-mel-22khz

Model Details

Model Description

Vocos is a fast neural vocoder designed to synthesize audio waveforms from acoustic features. Unlike other typical GAN-based vocoders, Vocos does not model audio samples in the time domain. Instead, it generates spectral coefficients, facilitating rapid audio reconstruction through inverse Fourier transform.

This version of vocos uses 80-bin mel spectrograms as acoustic features which are widespread in the TTS domain since the introduction of hifi-gan The goal of this model is to provide an alternative to hifi-gan that is faster and compatible with the acoustic output of several TTS models.

We are grateful with the authors for open sourcing the code allowing us to modify and train this version.

Intended Uses and limitations

The model is aimed to serve as a vocoder to synthesize audio waveforms from mel spectrograms. Is trained to generate speech and if is used in other audio domain is possible that the model won't produce high quality samples.

How to Get Started with the Model

Use the code below to get started with the model.

Installation

To use Vocos only in inference mode, install it using:

pip install git+https://github.com/langtech-bsc/vocos.git@matcha

Reconstruct audio from mel-spectrogram

import torch

from vocos import Vocos

vocos = Vocos.from_pretrained("BSC-LT/vocos-mel-22khz")

mel = torch.randn(1, 80, 256)  # B, C, T
audio = vocos.decode(mel)

Integrate with existing TTS models:

Matcha-TTS
Fastpitch

Copy-synthesis from a file:

import torchaudio

y, sr = torchaudio.load(YOUR_AUDIO_FILE)
if y.size(0) > 1:  # mix to mono
    y = y.mean(dim=0, keepdim=True)
y = torchaudio.functional.resample(y, orig_freq=sr, new_freq=22050)
y_hat = vocos(y)

Onnx

We also release a onnx version of the model, you can check in colab:

Training Details

Training Data

The model was trained on 4 speech datasets

Dataset	Language	Hours
LibriTTS-r	en	585
LJSpeech	en	24
Festcat	ca	22
OpenSLR69	ca	5

Training Procedure

The model was trained for 1.8M steps and 183 epochs with a batch size of 16 for stability. We used a Cosine scheduler with a initial learning rate of 5e-4. We also modified the mel spectrogram loss to use 128 bins and fmax of 11025 instead of the same input mel spectrogram.

Training Hyperparameters

initial_learning_rate: 5e-4
scheduler: cosine without warmup or restarts
mel_loss_coeff: 45
mrd_loss_coeff: 0.1
batch_size: 16
num_samples: 16384

Evaluation

Evaluation was done using the metrics on the original repo, after 183 epochs we achieve:

val_loss: 3.81
f1_score: 0.94
mel_loss: 0.25
periodicity_loss:0.132
pesq_score: 3.16
pitch_loss: 38.11
utmos_score: 3.27

Citation

If this code contributes to your research, please cite the work:

@article{siuzdak2023vocos,
  title={Vocos: Closing the gap between time-domain and Fourier-based neural vocoders for high-quality audio synthesis},
  author={Siuzdak, Hubert},
  journal={arXiv preprint arXiv:2306.00814},
  year={2023}
}

Additional information

Author

The Language Technologies Unit from Barcelona Supercomputing Center.

Contact

For further information, please send an email to langtech@bsc.es.

Copyright

License

Apache 2.0

Funding

This work has been promoted and financed by the Generalitat de Catalunya through the Aina project.

BSC-LT
/

vocos-mel-22khz