[META] On machine learning generative models
6mon 4d ago by sh.itjust.works/u/can in musicproduction@sh.itjust.works
I personally feel that goes against what this community should be about but I'd like feedback, especially on wording.
I don't feel generative "ai" audio is fitting here but a eurorack/modular/generative sequencer type thing would be welcome.
Given a recent post I believe the general community agrees but I need help with wording.
Any thoughts are appreciated.
Edit: not the focus, but I do think stem splitting models are fine, but I would interested in discussion about that too.
I think there’s a difference between say a sequencer that has a machine learning model behind it, and a model that’s trained on stolen music that takes words as input and outputs slop.
If you look at Synthesizer V for example, the voice providers were compensated and aware of what they are signing up for. The providers in the past back when it was concatenative synthesis who didn’t feel comfortable with a machine learning library, haven’t gotten one (e.g. Cangqiong). It’s also a much smaller company operating on a different scale. You won’t find Kanru Hua or Dreamtonics lobbying the American government bribing the U.S. president so they can be above the law. Sure, nowadays I’m sure they rent a server farm for their training, but I recall Hua training the initial models on his PC in the late 201Xs.
I think for me personally that’s where the sweet spot is, machine learning used to be a helpful tool in a creative process. Whether it replicates a voice, or an instrument, or helps filter noise. It can be done in a manner that doesn’t involve theft or the ruination of communities.
What do you think of Brian Eno's dabbling in generative music?
I'm not sure who he is or what he uses, so I can't say.
I don't oppose machine learning. It's a tool, and it's quite handy for things. For example, I make ample use of RNNoise If you were to say, record a bunch of samples from an instrument and train a model on that, I don't really see the problem with that.
My problem lies more in the underhanded tactics big corporations are doing right now, stealing masses of data, meddling hard in politics, destroying communities, etc. all for what's ultimately a rather underwhelming and extremely expensive end product no one even asked for.
You really should look him up. There's a good chance he's had a hand in, work that touched you, or at least inspired it. He pioneered ambient music for example.
See, this is something I wasn't aware of that further complicates this. I think I see things similarly but that makes it nkneasiwe to define a rule.
It makes it what?
lmao, sorry, I'm not even sure if I can parse that.
Essentially I really appreciate the input and nuance your comment provided. But despite that I have further difficulty in refining a comprehensive and specific "rule" for the community.
I could just enforce it based on "vibes" but it's my understanding that people don't like that approach to moderation. It may work now but ideally this place will grow and it could easily end up in prompt created audio spam territory.
Personally, I think "no gen ai music" would be well-enough understood as a rule to exclude AI generated songs but not music where AI has been used as a novel tool as part of the production process.
Alternatively, you could hold off on adding a rule like that and leave upvotes/downvotes as the method for evaluating posts with gen ai music. If this community starts to get inundated with low quality posts you could then revisit enforcing a rule?
Alternatively, you could hold off on adding a rule like that and leave upvotes/downvotes as the method for evaluating posts with gen ai music. If this community starts to get inundated with low quality posts you could then revisit enforcing a rule?
This could work for now. But with little traffic (admittedly partly my fault) it could still easily become a list mostly made up of downvoted slop.
"Thou shalt not make a machine in the likeness of a human mind."