Should dvector/autovc/wavenet-vocoder use the exactly same mel-spectrogram algorithm?

auspicious3000 / autovc

AutoVC: Zero-Shot Voice Style Transfer with Only Autoencoder Loss

https://arxiv.org/abs/1905.05879

MIT License

976 stars 207 forks source link

Should dvector/autovc/wavenet-vocoder use the exactly same mel-spectrogram algorithm? #70

Open KevinHua opened 3 years ago

KevinHua commented 3 years ago

The generated wave sounds are just noise after training and synthensis using default alogrithm provided in each project. I wonder which project's melsp algorithm should be used when processing, training and synthensis?

ruclion commented 3 years ago

I think each model can use different mels algorithm

the speaker encoder use it's mels as author use
autoVC decoder's training use the best hparams mels
vocoder's hparams just ensure the same with decoder's mels, so that it can become right waveform~