The most interesting thing here is the note at the bottom regarding computational cost: "A recent macbook pro reaches about 5 samples per second."
This shows how far this model is from realtime usage. However I'm sure Deepmind researchers are already looking into how to make this blockbased or some other optimization strategy.
And I should add that this was measured using a downsized model (just two blocks of dilated convolutions and a sampling rate of 4khz). Deepmind's paper does not report how many stacks are used to generate the samples, but I assume it's quite a bit more.
This implementation says: “A Tesla K80 needs around ~4 minutes for generating a second of audio at a sampling rate of 4000hz”, which is significantly faster.
90 minutes for 1s of audio was reported by someone from Google on twitter, but the tweet has been deleted.
I've clarified in the readme that my measurements are for a much lighter/smaller model than Deepmind's :).
Can someone elaborate on the usefulness of this implementation for Text-to-Speech?
I'm keen to experiment with voice synthesis. I want to create dialog, from multiple voice sources, for some characters in a VR application that I'm working on.
I guess I could do with an ELI5 on how I'd approach this with either of these libraries. I'm not familiar with any deep learning frameworks. But I am pretty handy with Python and have implemented SciKit stuff.
Also thinking this will give me a reason to try Azure K80 instance vs the AWS GPU instances I've been using for other stuff. That said, is a Tesla K80 the only option for WaveNet? I'm guessing I could run it on other GPU's but had read that memory might be an issue on some cards. If so what the lowest card I can run it on and will one of the AWS GPU instances suffice? I also have a GTX 970 at home, but I'm guessing that won't cut it.
If you have phonetically rich source data, festival will work pretty well. If you need a little more flex in your system, and can deal with a super weird training process, HTS is probably a better choice. With a small amount of work, you can use consolidated HTS models from within festival. (http://hts.sp.nitech.ac.jp/)
Further, if you pine for the fjords of DNN-land, merlin (https://github.com/CSTR-Edinburgh/merlin) is brand new and looking to make things a little easier for everybody.
But does anyone know if it's possible to do TTS with the recently released libraries?
Thanks for the links, but to my ear the samples on those links don't hit the mark. The Wavenet samples in the original article cross the threshold for me. So I'd like to try some short length dialog tests, especially as I've read elsewhere that 1 second only takes 4 minutes on a K80.
Any light anyone else can shed on this would be great.
Afaik none of the released libraries support the TTS experiment described in the paper. Deepmind used pre-computed linguistic features to guide the system in generating natural sounding speech, so your output will probably depend on the quality of those features.
For the sake of not spreading misinformation; the 4 minutes was measured using a small model with a sampling rate of 4khz, this would not generate something sounding like the samples from Deepmind.
In the paper, they say that they double the dilation factor up to a limit and then repeat:
1, 2, 4, ..., 512, 1, 2, 4, ..., 512, 1, 2, 4, ..., 512
The doubling of the dilation factor makes sense to me, but what is happening with the "repeat" part? I don't understand what they are trying go do. Wouldn't make more sense to continue doubling?
My intuition is that the doubling up to 512 does increase the receptive field, but you're essentially building a non-linear convolutional filter with a kernel size of 1024. The network benefits from stacking multiple of these groups, because each group can again convolve over the previous outputs at every temporal distance, which allows for learning deeper/higher level features. It is similar to the stacked 2d convolutions used for images, where every subsequent convolutional layers learns more abstract and higher level features/attributes of the data. This is just intuition though, there is no evidence yet that this holds for wavenet's architecture.
This shows how far this model is from realtime usage. However I'm sure Deepmind researchers are already looking into how to make this blockbased or some other optimization strategy.