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

How to generate speech from features with WORLD vocoder #29

Closed ghostcow closed 6 years ago

ghostcow commented 6 years ago

Hi,

I'm trying to debug my system that uses you WaveNet vocoder. Is there any way to create WAV from the features your code generates?

Thanks

kan-bayashi commented 6 years ago

What does it mean? You can run one of the recipes, then you can get generated wav file from feature vectors.

ghostcow commented 6 years ago

I meant, by using the WORLD vocoder instead of the WaveNet. My question essentially is, why do you change the F0 when creating the features? Is that really necessary?

kan-bayashi commented 6 years ago

I have tried both (f0 and continuous f0 + uv information), but I did not compare them strictly. The reason why I use continuous f0 is based on our experiences. Anyway, both work, therefore, maybe you can use f0 directly, instead of continuous f0 + uv.

And our feature extraction step is based on the sprocket, which use world in internal processing. By using sprocket.speech.Synthesizer, you can generate voice with world vocoder.

ghostcow commented 6 years ago

Thanks! I'll try your suggestions!