Neural vocoder support

Let’s discuss neural vocoder support.

1. Types

• Parallel WaveGAN (PWG), MelGAN I can easily implement these two since I am familiar with it. Kan-bayashi has packed his repo so we can simply pip install -U parallel_wavegan.
• WaveNet vocoder I don’t really prefer this one due to its slow inference speed.
• WaveGlow I don’t really prefer this one due to its slow training speed.

1. Usage & Structure We can add a synthesis stage to the recipe. We can provide pretrained models for users to download, and use an argument like voc_expdir to load the pretrained model. In addition, with PWG, kan-bayashi has also packed training code in the package, so we can provide recipes for users to train their own vocoders if they want. One example design can be like egs/pwg/vcc2018.

[unilight]

I see. I will train vocoders in kan-bayashi/ParallelWaveGAN and just provide pretrained model links in this repo. I will work on it next.

[k2kobayashi]

It’s nice to integrate PWG/MelGAN. Actually, I have already discuss it with @kan-bayashi. He said he can create recipe and pre-trained model after releasing vcc2020 dataset.

For the structure, I think following is nice.

• add stage 6 in egs/vaevc/template/run.sh
• implement egs/vaevc/<recipe>/local/download_pretrained_neuralvocoder.sh to download pre-trained models.
• implement crank/bin/generate_wav_{pwg,melgan}.py to generate wav file w/ pre-trained model and generated h5 in stage 5.