Question about duration loss

jaywalnut310 / glow-tts

A Generative Flow for Text-to-Speech via Monotonic Alignment Search

MIT License

660 stars 151 forks source link

Question about duration loss #55

Open Wangzhen-kris opened 3 years ago

Wangzhen-kris commented 3 years ago

Hi, I trained the model with my own dataset. However, the duration loss increases gradually after a period of time for val data, it's overfitting, right? Can I stop the duration prediction after several epochs?

Selimonder commented 3 years ago

Yeah, also in my experience duration predictor usually overfits quite fast. However beware of your validation set, if you have any abruptly stopped sentences.

Adding onto this question; Did anyone experimented with replacing duration predictor with FastSpeech2's?