Sunday, August 18, 2019

deep fake for making music?


So, I understand that deep fake can be used for guiding audio, not just video, and I now understand that one threat could be using deep fakes being sent to first responders.

I did search for 'deepfake music' but so far nothing.

My question is whether deep fake could be applied to music, so maybe we take one vocalist and map it over another, so Geddy Lee from Rush ends up singing Ice Cream Man by Van Halen. Not that I'm dying to do that, but it opens up some interesting possibilities. 

So, for example, separate the tracks that make up a song by Van Halen, and substitute the stylings of Rush. So, for example, have AI 'listen' to the entire body of work by a particular drummer, then it can play in that style.

I think there would be a break down at the point of writing lyrics unless it's open to non-sequitur, which would probably be my approach - leave it to people's imagination and let it be poetic without concrete meaning, kind of mad libs rule engine. 

So, have Bach writing melodies to Neil's drumming so it stretches and blends more like a rock song. Now you have an original instrumental composition, just add lyrics, then add vocals and bass. How would Geddy's bassline fit in? Well, take Bach's classical rules that would drive a more proper bassline, but add in a rule engine for Geddy's style. Layer on a funk filter. Swap in a Steve Harris bassline from Iron Maiden. Adjust to fit a particular period, like more 80s Rush, or more 90s Rush. Not that we really want this necessarily, but it's a starting point. Ultimately, don't we want originality? Genuine emotion? Isn't that what art is? Can a machine do that? Music today is all about style, not substance. It's formulaic as hell.

What would also be really nice for me - I want to discover new music. Forget about Pandora and Spotify. Let's have AI identify the patterns of music I like, then find music of a similar ilk. I actually thought about this in 1999, the idea was to be able to visually search music. The mood and/or composition would be represented by a color, like a heat map, and you could zoom into various colors to quickly identify variants within that range, filtering by attribute.

Of course, the extension of this would be to create music of an ilk that I like. But not necessarily just by copying styles as I described previously and tracking them over other compositions. Let's zoom into why of I liking certain music. It's not just the rhythm or the instrumentation or compositions or mood. That's what makes this question of creativity and originality interesting - what separates humans from AI. We're talking about personality and point of view. Philosophy and context and emotion, not merely style.