fatchord / WaveRNN

WaveRNN Vocoder + TTS

Home Page:https://fatchord.github.io/model_outputs/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Converting generated MEL spectrograms to Tacotron 2 format, in order to use a different vocoder?

serg06 opened this issue · comments

The WaveRNN vocoder is too slow for me (I'm running it on the CPU), so I want to hook up a different vocoder to the Tacotron for inferences.

MelGAN and HiFi-GAN are both very fast vocoders, and they both have pre-trained Tacotron2 models which I can use with Tacotron2 mel spectrograms without issues, but when I try inference Tacotron1 mels, the output is completely off.

It looks like converting Tacotron 2 mels -> Tacotron 1 mels is done like so or like so, so is there maybe some way to reverse this process to get Tacotron 2 mels?