Tacotron2: Plan to integrate with nv-wavenet?

[neverjoe]

I think integrating taco2 with nv-wavenet still is a long way, first thing i think nv-wavenet should support mol, not just 8bit.

[rafaelvalle]

Integration with nv-wavenet, including a PyTorch wrapper for nv-wavenet are ready and will be released soon.
Wavenet with Mixture of Logistics and Fine/Coarse outputs are being considered.

[neverjoe]

Great work!
Can u share some wav sample of your full stack tts system(8 bit nv-wavenet)?

[rafaelvalle]

Not yet but soon. We're currently working on another release.

[glennglennglenn]

@rafaelvalle How long do you think it will be before the integration with nv-wavenet is released? How is the voice quality with that integration?

[rafaelvalle]

Pytorch bindings for 8-bit mulaw nv-wavenet should be out this week.

[neverjoe]

nice work!@rafaelvalle, Can I use r9y9's wavenet arch of 8-bit mulaw with your work?

[rafaelvalle]

It should be possible. You'll need to load the weights onto the PyTorch bindings.

[neverjoe]

Nice, i will test it next week when i get my P40 GPU.

Rafael Valle [e-mail]于2018年5月14日周一 下午9:10写道：

It should be possible. You'll need to load the weights onto the PyTorch
bindings.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
https://github.com/NVIDIA/tacotron2/issues/3#issuecomment-388810628, or mute
the thread
https://github.com/notifications/unsubscribe-auth/ABsFmLy79Ngp0W0YRYKNwYSU-Fx80hMTks5tyYIsgaJpZM4TyB2X
.

>

Sent by Inbox

[glennglennglenn]

@rafaelvalle are the Pytorch bindings for 8-bit mulaw nv-wavenet going to be integrated into this repo? Or will we need to figure out how to get them to work together? I'm very excited to try this program and hear the results.

[rafaelvalle]

Integration is simple: just provide mel-spectrograms as input to the nv-wavenet pytorch bindings.
The pull request is here:
https://github.com/NVIDIA/nv-wavenet/pull/7/files

[rafaelvalle]

For a follow-up on this issue please use this https://github.com/NVIDIA/nv-wavenet/pull/7/files where the pytorch bindings are being released.
Closing.

[PetrochukM]

@rafaelvalle This is awesome. I've reviewed the README for the pull request, I have a couple questions:

• embedding_prev and embedding_cur, these are used for global conditioning via embeddings like speaker embeddings? Why are there previous and current embeddings?
• cond_input Is where we would locally condition the mel-spectrogram, right?
• Does this PyTorch module allow for training or just inference? If it does not allow for training, what is the recommended method for training?
• For the initial 2x1 casual convolution, where can we set the weight matrix?

[rafaelvalle]

@PetrochukM please ask these question on the nv-wavenet repo.

[PetrochukM]

@rafaelvalle Reposted!

[viogoes]

@rafaelvalle It should be possible. You'll need to load the weights onto the PyTorch bindings.
How to load the weights?

[rafaelvalle]

@viogoes please ask this question on the nv-wavenet repo.