WaveRNN Training Error

Immabot · January 28, 2020, 11:38am

Hey @erogol and TTS community
I’m trying to train the WaveRNN model for Tacotron2 which I’ve already trained
and i’m getting the following error.I cannot generate mel-spectrograms from the Tacotron2 model.
I’m getting the following error-

I think the dataset loader isn’t being created properly.Can you explain what the paths in below figures are supposed to be.

I’d appreciate any sorts of help on this.

Thanks!

Mike

erogol · January 29, 2020, 10:48am

can you try the notebook on the latest dev branch? If it does not work either, then there is a bug.

Immabot · January 30, 2020, 10:03am

Hey @erogol ,It Worked.Thanks.Appreciate it.
But I encountered another problem while training the vocoder.

I checked the WaveRNN/config.json file.I didn’t find a variable named mulaw.

Can you help solve this?
Thanks!

erogol · January 30, 2020, 10:05am

you can add mulaw field. It is just a json in the end.

Immabot · January 30, 2020, 10:16am

Hey @erogol,
Actually there are 3 fields-
mulaw,use_aux_net and use_upsample_net.
What are it’s values supposed to be?

erogol · January 30, 2020, 10:17am

can you link me the branch you use ?

Immabot · January 30, 2020, 10:18am

Yeah.

erogol · January 30, 2020, 10:18am

this might help https://github.com/erogol/WaveRNN/blob/master/config_libri360.json

erogol · January 30, 2020, 10:20am

otherwise, if you use auxnet, set use_aux_net=true and use_upsample_net likewise.

Immabot · January 30, 2020, 10:20am

Okay.
Thanks again.
Appreciate it.