How to prepare TTS-synthesized mel-spectrograms for vocoder training?

[dimitry12]

Given a custom trained TTS, I want to train a vocoder. I want to use teacher-forced mel-spectrograms generated by TTS for vocoder training, as they do in Tacotron2 paper or as @erogol does in #26 (comment)

First, is there a script or a notebook for generating teacher-forced mel-spectrograms given a TTS model?

Second, once these mel-spectrograms are generated, is

TTS/TTS/vocoder/configs/wavegrad_libritts.json

Line 64 in c657124

"feature_path": null, // if you use precomputed features

the right place to point vocoder to it?

[lexkoro]

First, is there a script or a notebook for generating teacher-forced mel-spectrograms given a TTS model?

https://github.com/mozilla/TTS/blob/master/notebooks/ExtractTTSpectrogram.ipynb should be a good start

Second, once these mel-spectrograms are generated, is the right place to point vocoder to it?

Yes that should be correct

[dimitry12]

@sanjaesc thank you!