Differentiable DSP

P

plexi59

Guest
So the other day I mentioned that it's a bit too early to be talking about truly "neural" DSP and we're nowhere close to it yet. Apparently, though, at least someone at Google is working on it: https://storage.googleapis.com/ddsp/index.html. Check out their voice to violin tone transfer especially, since it's closer than other examples to what Fractal Audio aims to do. It's in very, very early stages, but give it time and we could end up with something interesting. Still nowhere near realtime for either training or inference, and even if it is eventually made realtime, the latency most definitely won't be in single digit milliseconds.

To those inclined to dismiss this out of hand, recall that speech synthesis systems were laughable even just a few years ago, and look at them now, most people can't tell them from a real human in a phone call anymore.
 
Back
Top Bottom