About WaveRNN, not Alternative. about wavernn

[bjtommychen]

like this:
input = torch.cat((prev_y, mels), dim=1)
coarse_input_proj = self.I_coarse(input)
...
fine_input = torch.cat([prev_y, mels, current_coarse], dim=1)
...

[fatchord]

@bjtommychen I haven't conditioned the first three notebooks on spectrograms. Did you implement conditioning yourself?

[bjtommychen]

Yes. I did. after more training, quality become better, and still worse than Alternative mode.

[kubicndmr]

@bjtommychen can you show how you conditioned WaveRNN by yourself

[hdmjdp]

@bjtommychen same question

[hdmjdp]

@bjtommychen 3q， I just try to implement the origin rnn in Alternative model. But I can not got a better quality than Alternative one.

[bjtommychen]

@hdmjdp same as me. the origin model's quality is worse than Alternative one. and also training slower.

[ohleo]

@bjtommychen
Could you share the synthesized sample of 16-bit WaveRNN and training sequence length(=seq_len in notebook3)?
I think training sequence length should be 7936 samples(0.359 sec in 22.05kHz data) like r9y9 wavenet code. But, Training-time will be unrealistic.
How many training step is needed for intelligible sound?

I would appreciate your reply.