Parallel Wavenet-Vocoder

[twidddj]

Planed TODO

• KL + Power - Single speaker

Properties not specified in the paper

• Sampling number for the loss (We may have some limitation for GPU)
• Number of mixture for IAF layers
• Averaging method for Power loss
  • ex) Just reduce_mean on time axis or using moving average or ..
• .. (Please, let us know those)

Another implementations

• https://github.com/zhf459/P_wavenet_vocoder (used r9y9's wavenet in pytorch)

[neverjoe]

sadly, there are many details behind the paper, i find nobody can reproduce result.

[twidddj]

Most of all, I'm not sure mel-spectrogram would fit well to reproduce result like the linguistic features in the paper. We may have to consider another constraints to make up for the weakness of mel-features.

[neverjoe]

I do not think mel-features is the key problem of me, i think the iaf and probability Density Distillation is very very import for the quality.

[maozhiqiang]

@neverjoe do you tell me how to connect teacher model to eval student !Are they all trained in one sess! thanks

[neverjoe]

U just need to read the paper, the paper tell u clearly. maozhiqiang <[email protected]>于2018年4月26日周四 上午10:44写道：

@neverjoe <https://github.com/neverjoe> do you tell me how to connect teacher model to eval student !Are they all trained in one sessiotion! thanks — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABsFmPaWAZz6iZtoo8y2Uhc1fpbsAUFUks5tsTSSgaJpZM4Te5J2> .

-- Sent by Inbox

[maozhiqiang]

hi ! How to let teachers' network parameters do not participate in training? thanks

[xuerq]

@maozhiqiang
"tf.stop_gradient" in tensorflow

[maozhiqiang]

thank you @xuerq

[neverjoe]

@xuerq have u got quite quality wav by parallel wavenet？ maozhiqiang <[email protected]>于2018年4月29日 周日上午9:24写道：

thank you @xuerq <https://github.com/xuerq> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#5 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABsFmIqA5wI8f3bW5WHNCv3iewpAuHEuks5ttRY8gaJpZM4Te5J2> .

-- Sent by Inbox

[xuerq]

@neverjoe still working on it，the wav sampled from student is not so good as teacher

[maozhiqiang]

@xuerq ,@neverjoe @twidddj how to use teacher model to eval students output? Is it use training process to assessment or using generative process to assessment? thanks!

[maozhiqiang]

@twidddj do you get reasonable result about paralle wavenet!

[twidddj]

Hi @maozhiqiang, we are still trying to get better result of it. We have got some results, but It's not enough. I attached the results here. Thanks for your attention of our project!

[maozhiqiang]

hi @twidddj ! Thank you very much for your reply！　do you used KL loss?