Better Quality?

[PetrochukM]

Samples this repo sound too real to be true. What's the difference between these repos?

https://github.com/kan-bayashi/PytorchWaveNetVocoder

[bpotard]

The acoustic conditioning features are different; r9y9 uses 80 mel spectrogram features with a 62ms step, while kan-bayashi uses 28 acoustic (mcep+f0+noise) features with a 5ms step. Also, the wavenet setup is a bit different, if I am not mistaken it seems kan-bayashi uses the equivalent of: "layers": 30, "stacks": 3, "kernel_size": 2

So kan-bayashi use more acoustic features for input conditioning (a smaller vector for each frame, so a lower spectral precision but a much higher sampling rate, so a better time precision) and has a slightly larger and deeper wavenet. I believe the higher sampling rate in input conditioning explains most of the difference in acoustic quality.

But you can reproduce this setup fairly easily with https://github.com/r9y9/wavenet_vocoder and using WORLD for the acoustic feature extraction.

From my own experiments, the vocoding sounds better with the wavenet setup above, but it requires a lot of memory and time to train.

[r9y9]

Anther difference is that they use noise shaping technique to reduce quantization noise, while we don't.

You can find paper abstract there: Mel-cepstrum based quantization noise shaping applied to speech synthesis based on WaveNet.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.