patriotyk's picture
Update
77aded6 verified
|
raw
history blame
2.15 kB
---
license: mit
pipeline_tag: text-to-speech
tags:
- vocos
- hifigan
- tts
- melspectrogram
- vocoder
- mel
---
### Model Description
<!-- Provide a longer summary of what this model is. -->
**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](https://github.com/jik876/hifi-gan/blob/master/meldataset.py)
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.
### Installation
To use Vocos only in inference mode, install it using:
```bash
pip install git+https://github.com/langtech-bsc/vocos.git@matcha
```
### Reconstruct audio from mel-spectrogram
```python
import torch
from vocos import Vocos
vocos = Vocos.from_pretrained("patriotyk/vocos-mel-hifigan-compat-44100khz")
mel = torch.randn(1, 80, 256) # B, C, T
audio = vocos.decode(mel)
```
### Training Data
The model was trained on private 800+ hours dataset, made from Ukrainian audio books, using [narizaka](https://github.com/patriotyk/narizaka) tool.
### Training Procedure
The model was trained for 2.0M steps and 210 epochs with a batch size of 20. We used a Cosine scheduler with a initial learning rate of 3e-4.
We where using two RTX-3090 video cards for training, and it took about one month of continuous training.
#### Training Hyperparameters
* initial_learning_rate: 3e-4
* scheduler: cosine without warmup or restarts
* mel_loss_coeff: 45
* mrd_loss_coeff: 1.0
* batch_size: 20
* num_samples: 32768