Espnet: Question about Learning curves of FastSpeech.

[dathudeptrai]

Do u have any learning curves (tensorboard) of fastspeech on LJSpeech dataset ?, i cann't find it in this repo.

[kan-bayashi]

Hi @dathudeptrai. I will share the training curve figure.

This is the training curve of the following model (fastspeech.v4 + taco.v3 as teacher)
https://github.com/espnet/espnet/blob/master/egs/ljspeech/tts1/RESULTS.md#models-1

[dathudeptrai]

hi @kan-bayashi seem the validation loss so high, i have similar behavior on tensorflow model now :D. BTW, l1_loss here is just mae of mel grountruth and mel_pred ?

[kan-bayashi]

Right. FastSpeech has the tendency to overfit in my experiments. This model used knowledge distillation, so the target is not ground-truth but the Mel predicted by the teacher model without teacher forcing.

[dathudeptrai]

hi @kan-bayashi , got it. I will try some new idea to avoid overfit problem now :D, i will share with u when the idea work :D.