3 ms·

I can assure you there is a lot of interest in the Machine Learning field to make deep fakes for music. The problem is deep fakes are not time-series based. Rig

I can assure you there is a lot of interest in the Machine Learning field to make deep fakes for music. The problem is deep fakes are not time-series based. Right now a video is treated as a bunch of independent images, whereas you couldn't do that for music/text as there are long term dependencies between time steps. More specifically, there is a recent paper from ICLR about audio GANs (I haven't read it) https://openreview.net/forum?id=H1xQVn09FX https://openreview.net/forum?id=H1xQVn09FX. IMO, this tech should come out (even in app form) in the next 2-4 years.

hirundo 8y ago
The coming age of infinite mashups. Bruce Springsteen's Fire sung by Elmer Fudd, to infinity. What does a cross between Elvis and John Philip Sousa sound like? Does Jimi Hendrix need a little more Abba? Such an ML generator could explore a space of musical styles with more nodes than there are quarks. Now randomly upload a few to YouTube, let upvotes be the fitness test to select the next generation, and repeat forever. Could this algorithm top the charts or would it devolve into random crap?