fast speech mel normalization

[superhg2012]

when compute mean and scaler for mel-spectrogram before normization, mean and scaler are computed from all dataset and only the first frame mel?

mel = mel[0].numpy()

[dathudeptrai]

@superhg2012 mean, var for all TRAINING SET. shape mel = [1, len, 80] so i use mel[0]

[superhg2012]

All right, I get it, thanks! By the way replace relu activation in ffn layer of FFT block with mish activation function really works for me.

[dathudeptrai]

@superhg2012 Fastspeech without duration from teacher model is comming :))) let see

[superhg2012]

@superhg2012 Fastspeech without duration from teacher model is comming :))) let see

any paper reference?

[dathudeptrai]

@superhg2012 flow-tts, glow-tts

[dathudeptrai]

@superhg2012 i just added fastspeech samples, can you take a look and see if it's good or not :D. https://dathudeptrai.github.io/TensorflowTTS/