search

r9y9 / wavenet_vocoder

WaveNet vocoder
https://r9y9.github.io/wavenet_vocoder/
Other
2.31k stars 499 forks source link

"quantize_channels" parameter in JSON file #163

Closed byuns9334 closed 5 years ago

byuns9334 commented 5 years ago

Thanks for the great work!

If I use the 25-bit audio samples for training data, should I set "quantize_channels" parameter as 2^15=33554432? if not, how can I set this parameter?

r9y9 commented 5 years ago

I guess you might want to learn how parallel wavenet deals with the problem. See sec. 2.1 of https://arxiv.org/abs/1711.10433 for details.

In short: we can consider using an explicit parametrized distribution to model raw audio.