keonlee9420 / robust_fine_grained_prosody_control Goto Github PK

Hi, I'm implementing this paper recently as well :). I have some confusion about your code:
(1) There is no ReLu layer after the prosody embedding, which is used in the original paper.
(2) You adopt LSA (which I think is not robust). Have you run into the problem that LSA fails frequently when the prosody is concatenated to the text encoder output?
(3) Could you provide some listening demos?

Hello and thank you for sharing your work!
Do you have a pretrained model to share?
Thank you very much in advance!

maybe , it is -1e9?

안녕하세요, LibriTTS 100시간으로 모델을 학습시키려고 하는데, 96000 step을 진행한 이후에 tensorboard를 통해 attention alignment를 확인해보았지만 attention alignment가 diagonal하게 잡히지 않아서 혹시 코드에 추가적으로 진행해줘야 attention alignment가 diagonal하게 나오는지 문의드립니다.