barronalex / Tacotron

Implementation of Google's Tacotron in TensorFlow
236 stars 80 forks source link

reshape_frames before logarithm not after #3

Closed candlewill closed 7 years ago

candlewill commented 7 years ago

As reshape_frames pad 0 to input, we should first do reshape_frames, then logarithm. Otherwise, we should pad with -inf if do logarithm first.

barronalex commented 7 years ago

You're absolutely right but I ended up entirely rewriting the padding code so that silence would be correctly reconstructed. Very sorry not to have fixed this earlier.