Finetuning Tacotron2 on your pretrained model about thorsten-voice HOT 6 CLOSED

Thanks for your nice feedback on my work @Mareike-RTY .
Before you struggle on finetuning with existing Thorsten TTS-models maybe you want to wait some days until the new TTS-models will be released. We're in the final phase before release.

The new models will provide a much better and natural speech flow and for that might be a better startingpoint for your finetuning process.

Here you can hear the difference between "current" and "new" Thorsten model.
https://www.thorsten-voice.de/2022/03/20/vergleich-thorsten-aktuell-mit-dem-neuen-modell/

from thorsten-voice.

in order to hopefully solve the unpickling error mentioned above, maybe you could let me know how you pickled the VITS model and which Python version was used? I have read a post saying "if the model is not unpickled the same way it was pickled, unpickling cannot work".

@Mareike-RTY The model was trained in a Python 3.6 environment.

from thorsten-voice.

Thanks for your reply, @thorstenMueller! Then I will gladly wait until the new models are released. I'm already very excited - the "new" Thorsten model sounds highly natural! Wow!

from thorsten-voice.

Dear Thorsten,

I have just tested your newly released VITS model on the TTS server, and the generated audios really sound amazing! Great work, thank you.
I found out that the new model is in the same format (data.pkl) as the previous one, so unfortunately, I cannot resolve my original thread on my own because I do not know how to use this kind of model instead of a .h5 model for finetuning.

Would you please be so kind to answer my original question and provide me maybe with a little information of how I could proceed? I have tried it with several installations of torch, so this is probably not the issue. And I don't load the model multiple times but only once.
Thank you very much in advance!

Have a nice weekend.

from thorsten-voice.

Dear Thorsten,

in order to hopefully solve the unpickling error mentioned above, maybe you could let me know how you pickled the VITS model and which Python version was used? I have read a post saying "if the model is not unpickled the same way it was pickled, unpickling cannot work".

Thank you very much, and have a nice day.

from thorsten-voice.

Update: In the meantime I started using coquiTTS instead of TensorFlowTTS for finetuning the VITS model.
Now it's working, so I guess this issue can be closed.

from thorsten-voice.