BSC-LT / wavenext-mel

Wavenext-mel-22khz

Model Details

Model Description

Wavenext is a modification of Vocos, where the last ISTFT layer is replaced with a a trainable linear layer that can directly predict speech waveform samples.

This version of Wavenext 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.

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.

Usage

Installation

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

pip install git+https://github.com/langtech-bsc/wavenext_pytorch

Reconstruct audio from mel-spectrogram

import torch

from vocos import Vocos

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

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

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)

Training Details

Training Data

The model was trained on 4 speech datasets

Training Procedure

The model was trained for 1M steps and 96 epochs with a batch size of 16 for stability. We used a Cosine scheduler with a initial learning rate of 1e-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: 1e-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 96 epochs we achieve:

• val_loss: 3.79
• f1_score: 0.94
• mel_loss: 0.27
• periodicity_loss:0.128
• pesq_score: 3.27
• pitch_loss: 31.33
• utmos_score: 3.20

Citation

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

@INPROCEEDINGS{10389765,
  author={Okamoto, Takuma and Yamashita, Haruki and Ohtani, Yamato and Toda, Tomoki and Kawai, Hisashi},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)}, 
  title={WaveNeXt: ConvNeXt-Based Fast Neural Vocoder Without ISTFT layer}, 
  year={2023},
  volume={},
  number={},
  pages={1-8},
  keywords={Fourier transforms;Vocoders;Conferences;Automatic speech recognition;ConvNext;end-to-end text-to-speech;linear layer-based upsampling;neural vocoder;Vocos},
  doi={10.1109/ASRU57964.2023.10389765}}

@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 [email protected] .

Copyright

Copyright(c) 2024 by Language Technologies Unit, Barcelona Supercomputing Center.

License

Apache 2.0

Funding

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