no mel-spectrogram saved in preprocessing stage

begeekmyfriend / tacotron2

Forked from NVIDIA/tacotron2 and merged with Rayhane-mamah/Tacotron-2

BSD 3-Clause "New" or "Revised" License

81 stars 38 forks source link

no mel-spectrogram saved in preprocessing stage #21

Closed MorganCZY closed 4 years ago

MorganCZY commented 4 years ago

It seems you forget to save mel-spectrogram to the disk. Should I directly save "mel_spectrogram" or "mel_spectrogram.T"?

begeekmyfriend commented 4 years ago

This comment line would save your disk space and the ground truth Mel spectrograms are useless in training. You can generate GTA Mel spectrograms with scripts/gta_synth.sh.

MorganCZY commented 4 years ago

Why are the ground truth Mel spectrograms useless in training? Based on my understanding, tacotron2 is trained to learn a map between text sequence and mel. Can the training work if giving only texts but no corresponding the true mel?

begeekmyfriend commented 4 years ago

Sorry, I said wrong...The ground truth Mel spectrograms would be feed during data loading in place. Therefore it no need to save as numpy files.

MorganCZY commented 4 years ago

I just checked codes about the class TextMelDataset. I now understand your design.