search

kan-bayashi / PytorchWaveNetVocoder

WaveNet-Vocoder implementation with pytorch.
https://kan-bayashi.github.io/WaveNetVocoderSamples/
Apache License 2.0
297 stars 57 forks source link

Is adaptation of wavenet vocoder possible? #43

Closed cdemir1919 closed 5 years ago

cdemir1919 commented 5 years ago

Thank you for your work. is it possible to train wavenet vocoder with multi-speaker data and adapt it to a target speaker with limited data? if yes, how can i do that?

kan-bayashi commented 5 years ago

Yes, it can be done. We often do it in the scenario of voice conversion. For example, the recipe of multi-speaker WaveNet vocoder is available in egs/arctic/si-close of egs/arctic/si-open, you can train WaveNet vocoder using 6 speakers' data via the recipe. And then you can re-train the WaveNet using the multi-speaker WaveNet as the initial model. This can be done with --resume option in train.py. https://github.com/kan-bayashi/PytorchWaveNetVocoder/blob/86688e34510a9e2382b148a6c0f865c0c85ce215/src/bin/train.py#L496-L505 Maybe it is not necessary to load the states of optimizer, only that of model is OK.

cdemir1919 commented 5 years ago

Thank you for your explanation.