HiFi Gan: Spectograms in training

[harveenchadha]

Hi,

I want to know , how spectograms are generated from wavs in the training of HiFi GAN, shouldn't the spectogram come from spectogram generator like GlowTTS, Fastpitch for efficient training?

[erogol]

It might improve the results using generated specs but using real specs should also work well.

You can use TTS/bin/extract_tts_spectrograms.py if your model is glowTTS or Tacotron. Otherwise, you need to compute them manually.