search

xcmyz / FastSpeech

The Implementation of FastSpeech based on pytorch.
MIT License
858 stars 213 forks source link

Have anyone tried using LSTM to replace FFT block? #84

Open BuaaAlban opened 4 years ago

BuaaAlban commented 4 years ago

I have trained [37800/192000] steps, and it seems won't converge to a good value, especially the duration loss, it doesn't change much.

Mel Loss: 2.8434, Mel PostNet Loss: 2.5580, Duration Loss: 2.3693;