yistLin / universal-vocoder

A PyTorch implementation of the universal neural vocoder

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Universal Vocoder

This is a restructured and rewritten version of bshall/UniversalVocoding. The main difference here is that the model is turned into a TorchScript module during training and can be loaded for inferencing anywhere without Python dependencies.

Generate waveforms using pretrained models

Since the pretrained models were turned to TorchScript, you can load a trained model anywhere. Also you can generate multiple waveforms parallelly, e.g.

import torch

vocoder = torch.jit.load("vocoder.pt")

mels = [
    torch.randn(100, 80),
    torch.randn(200, 80),
    torch.randn(300, 80),
] # (length, mel_dim)

with torch.no_grad():
    wavs = vocoder.generate(mels)

Emperically, if you're using the default architecture, you can generate 30 samples at the same time on an GTX 1080 Ti.

Train from scratch

Multiple directories containing audio files can be processed at the same time, e.g.

python preprocess.py \
    VCTK-Corpus \
    LibriTTS/train-clean-100 \
    preprocessed # the output directory of preprocessed data

And train the model with the preprocessed data, e.g.

python train.py preprocessed

With the default settings, it would take around 12 hr to train to 100K steps on an RTX 2080 Ti.

References

About

A PyTorch implementation of the universal neural vocoder


Languages

Language:Python 100.0%