A TensorFlow Implementation of DeepMind's WaveNet Paper

https://github.com/ibab/tensorflow-wavenet

by ot • 10 years ago

185 30 10 years ago

alexbeloi• 10 years ago

I had plans to do the same, I'm glad somebody beat me to it.

Kenji• 10 years ago

That is an admirable mindset - I cannot help but be a bit frustrated when someone independently implements my ideas first. On the one hand, it validates the idea. On the other hand, building it would be fun, but now that someone else did it, doing it again would be akin to reinventing the wheel and it is more productive to turn your attention to something new (unless you think you can execute the project much better).

sherjilozair• 10 years ago

This project (and others in different frameworks) are far from complete. The prize for reproducing WaveNet in open source remains unclaimed.

vintermann• 10 years ago

In this case, it's DeepMind's idea anyway :)

I get more disappointed when the opposite happens. I think something like, "Yeah, I'm totally going to add support in torch for noisy activation functions like in this paper!" (https://arxiv.org/pdf/1603.00391.pdf). Then I procrastinate and put it off. Then I think, "No matter, someone else has surely done it by now". Then they haven't.

jonnycowboy• 10 years ago

Do you know how long it took to train using that dataset and with what hardware configuration?

Dowwie• 10 years ago

I wonder whether accents could be layered on top of the trained data?

aab0• 10 years ago

Probably. You can add in 'speaker' as a bit of metadata to the samples (this is what is meant by 'conditioning on') and teach it to speak like different people, so if you have a diverse sample of speakers and you add in 'accent' as another variable, it might well learn to disentangle individual speakers from their accents and then you can control generated accents by changing the metadata.

teddyknox• 10 years ago

Doing a forward pass for every sample sounds like it would be prohibitive for real-time applications.

nicklo• 10 years ago

It absolutely is. DeepMind reported that 1 second of audio generation takes about 90 minutes to generate.

throwawaymsft• 10 years ago

Assuming it's computation bound, it's a factor of 5400 (~13 doublings in CPU power required to get to real-time, assuming no algorithmic improvements).

Itsdijital• 10 years ago

Was that in the paper? I was looking for a source for it last night but couldn't come up with it

lallysingh• 10 years ago

Can we just use 90 cores?

c3534l• 10 years ago

We're still a couple of papers before we the computation down to a reasonable amount. Or eventually Moore's Law will take care of it. It might have applications that aren't real-time, too. I'm writing a video game in my spare time, and I was wondering how I would do the foley. If I could feed in some sounds and synthesize a library of sound effects that all sound different enough that it won't be repetitive to hear that same exact footstep sound for the entire game, then I consider that a win. So, you know, this is cutting edge research we're talking about.

blaurence5• 10 years ago

Here's a Theano implementation: https://github.com/huyouare/WaveNet-Theano

sargun• 10 years ago

I'm really interested in building an RNN and training it against something like movies. I'd love to then take it and input a song and translate it to a music video that's a composite. I'm also interested in the legal ramifications of doing such a thing...

Does anyone know of prior art?

A TensorFlow Implementation of DeepMind's WaveNet Paper

Related Stories