WaveRNN modifications?

[mrgloom]

Is this repo uses vanilla version of WaveRNN (https://github.com/fatchord/WaveRNN)? or architecture was modified and model was retrained?

It's quite fast relative to my benchmark of tacotron2 + wavernn from here https://github.com/erogol/WaveRNN but (maybe) have less natural voice.

BTW I don't tried new universal vocoder version.

[m-toman]

What I've seen the most significant differences are 16kHz sampling rate and a slightly smaller window size for the batched synthesis.

[mrgloom]

I wonder what is lower limit of sample_rate for reasonable speech quality? Can it be tested just by resampling by something like ffmpeg?

1. https://github.com/fatchord/WaveRNN
   sample_rate = 22050
   https://github.com/fatchord/WaveRNN/blob/master/hparams.py#L20

2. https://github.com/CorentinJ/Real-Time-Voice-Cloning
   sample_rate=16000
   https://github.com/CorentinJ/Real-Time-Voice-Cloning/blob/master/vocoder/hparams.py#L6
   https://github.com/CorentinJ/Real-Time-Voice-Cloning/blob/master/synthesizer/hparams.py#L117

3. https://github.com/erogol/WaveRNN
   mold model:
   "sample_rate": 22050
   https://drive.google.com/drive/folders/1wpPn3a0KQc6EYtKL0qOi4NqEmhML71Ve
   10bit model:
   "sample_rate": 22050
   https://drive.google.com/drive/folders/1VnTJfg2zmvochFNyX7oyUv9TFq6JsnVp
   universal vocoder:
   "sample_rate": 16000
   https://drive.google.com/drive/u/0/folders/15JhAbc91dT-RRZwakh_v4tBVOEuoOikg

[CorentinJ]

What I've seen the most significant differences are 16kHz sampling rate and a slightly smaller window size for the batched synthesis.

This is correct. I haven't changed anything else other than the data loader.