auspicious3000/SpeechSplit

Does it will work on unseen data also? will it be able to convert voice of unseen speaker with different content than that of data in training, will we obtain the disentanglement?

pycodebook opened this issue · 8 comments

pycodebook commented
Does it will work on unseen data also? will it be able to convert voice of unseen speaker with different content than that of data in training, will we obtain the disentanglement?
OSSome01 commented

I have the same doubt, can someone please clarify.
Thanks in advance.

  • 👍1
zhouyong64 commented

I have the same question.

auspicious3000 commented

You can make it generalize to unseen speakers by training it the same way as AutoVC.

skol101 commented

@auspicious3000 could you explain what you mean by 'training it the same way as AutoVC"?

Repeat all steps from here https://github.com/auspicious3000/autovc#2train-model ?

Or change make_metadata.py in SpeechSplit to embed speaker encodings but train using model from SpeechSplit?

auspicious3000 commented

@skol101 it means training with generalized speaker embeddings instead of one-hot embeddings

  • 👀1
skol101 commented

I did that -- used make_metadata.py from AutoVC.
Now I have removed validation part from the solver.py in this repo, because there's nothing to validate it against (as in solver_encoder.py in AutoVC) and started training.

Am I doing this correctly? Your help is very appreciated.

auspicious3000 commented

sounds correct, but you don't need to remove the validation part

  • 👍1
  • 👎1
  • 😄1
skol101 commented