16 comments

[ 3.2 ms ] story [ 39.4 ms ] thread
Very nice. Anyone know of projects that aren't tackling the full-song problem but rather instrument parts/loops/stems/acapellas? I'd like something that's more like "infinite AI Loopcloud/Splice" most of these full-song models don't do well to be asked for individual parts in my experience (though I will have to try it with this one).
This gets discussed a lot but unfortunately there's just not much out there around this.

The closest thing I've seen is virtual drummers in Logic X which will follow along with the structure of your song and generate a percussive accompaniment. It's no substitute for a real drummer but it's serviceable.

Honestly surprised no one's made a true "AI-Splice" yet
Something interesting... the first 10 seconds or so of the "Death Growl" example[1] is basically copied verbatim from "Ov Fire And The Void" by Behemoth.

More specifically, I think the part that seems copied is at 2:13 of the original[2], as it leads into a solo-ish bit which in the AI version sounds similar still, but goes on to do its own thing:

[1] https://map-yue.github.io/music/moon.death_metal.mp3

[2] https://youtu.be/vAmnsKKrt9w?t=133

Does Shazam think it is the same?
The youtube link is suddenly not available any more (at least in the UK)
Funny because since the Blurred Lines lawsuit you can be infringing for using the same chord progression.
I had many times on Suno when asking for something, for instance, a metal song with melodious guitar solos, it basically, almost note for note but NOT completely, Megadeth Marty Friedman, especially from around Rust In Peace times. It's good, but why does it pick that to copy specifically?
What is the use case for music generation models? I see usecases for alot of the other foundation models like text, image, tts, sst, but why do I want AI generated music?
yeah, but have yall made any progress in a model that can have sex with my partner for me?
Not using streaming services help somewhat, but I'm not looking forward to having to vet artists I come across for whether they did any substantial original thinking or work for themselves. Tired of snake oil foolishness.
did a fun experiment around this in 2015 based in karpathy’s unreasonable effectiveness of RNN’s post.

i kept plucking away at it until i got it to a point where it could generate sheet music and guitar tabs in the style of various artists.

would be fun to revisit that project with fresh eyes.

I get the incentives for full-song generation models, but it looks like it's the only thing that pops up. Where are the audio models I can use with a positive effect while working on music? Style transfer for instruments, restoration (but not just trained on over-compressed mp3s, talking about bad recording conditions), audio-to-audio editing?

You'd think those would be easier to achieve than something that tries to just replace me completely.