Want a sampler or an update for existing hardware sampler (DT, OT, MPC, Push 3 ) to instantly generate sounds and/or lyrics by a prompt with some tweaking (size, pitch, character, stability, age etc… ) and instantly use it on sampler’s track.
Do you think it would be possible in near future and what could be limitations for this?
No offense, but why?
Today, the device would need to be as powerful as a mid-to-high end mobile phone with A LOT of memory. The models against which the AIs run are massive data files and need to be fully in memory, and with a decent GPU available. And even then, they’ll run sloooow. Both models and processors will improve in the next few years so the hardware cost will go down.
why not?
to speed up things, open up imagination, maybe find out some new sounds never existed before.
Maybe not AI exactly but just a randomize patch button combined with a pattern generator that will input random notes in a random scale/tempo. Could be cool. Like auto fill in a tracker.
Now’s a good time to start a business building this. There’s a lot of money swishing around venture capital for interesting AI projects.
(although, luddite me suggests that if you have time to do that, you have also go time to find some friends who sing, and to collaborate rather than hole up on your own with your voice-bot)
Well when the AI vs human war comes we can buy ourselves some time by tricking the AI and bogging it down with making snares.
Another reason could be - copyright. Realeasing some music atm and a bit frustrated with finding owners of some samples I’ve used, and still not sure if I’ll be able to legalise it. With AI samples it won’t be a problem.
Another thing could be - you let it hear some music you have in mind and let it generate something similar but with another harmony or rhythm you need, like it could listen to external sample and compare it with what you have programmed in sequencer and generate something which will fit.
So many ideas…
Yeah, but I would like to generate some new sounds, not to emulate existing singing or playing style human can do… like it was with 303 - they built it for emulating electric bass but at the end it became completely different thing.
For example “Acid bassline made of human voice with some elephant scream accents” … aha… and add it to granular engine and see what it will sounds like…
I think tech like this is definitely going to be a big part of the future of sampling and synthesis in general, being able to imagine up any sound and have it be actually manifested will be huge. Natural language prompting of some kind of LLM to generate arbitrary sounds seems like a relatively easy first step towards this, I’m guessing there’s some stuff out there already doing something like this.
Um I don’t remember what it was specifically but @captain8 posted a topic about some really tedious looking sample generation tool. There was an attached video in the thread which I watched and though the results were interesting, it looked like way more work to me than just coming up with your own shit, then resampling it into a lofi sampler for character then bouncing it back and mangling it, and that shit is tedious, so when I say the other thing looked like more work I really mean it.
…or it could be a connected device to let the processing happen elsewhere.
I don’t believe you can’t do it like that now.
AI’s might make it happen faster and/or allow you to request imaginary signal flows which are currently out of your reach.
But… the thing is… at the moment, these models are based on existing music. So, at best, they’ll be the same as what you can get now by working with existing gear and people. Not “copies of” but “inspired by”. Humans are still, currently, better at making up new useful stuff.
I think it’ll get more powerful when AI’s are able to write novel algorithms and engines in realtime in response to suggestions, rather than by generating “inspired by” sounds carved out of white noise like they do today. We’re an away off doing unmonitored full application development. Most of the code AI generates now needs manual reshaping to be useful.
I mean… I’ve heard a few full song generators that are amazing in terms of audio quality, but … they sound like people doing stuff you’ve heard before. And I’ve heard stem generators that are basically churning out the equivalent of sample CDs. I don’t research this too deep because the whole field scares me (what do I do with my gear now it’s redundant?), so feel free to post better suggestions.
Are you at all aware of how these AI models are trained? They utilize vast amounts of existing content. The copyright implications of that are so murky at present they make sample clearance look crystal clear by comparison.
Yeah, but again, these AI models are trained to replicate trends in an existing corpus of content, and relate those trends to specific words. They’re not going to innovate new sounds for you, only recombine old ones.
An LLM is great for imitating the status quo. If you want truly new sounds, you’re better off getting a copy of Metasynth and painting the sound spectra visually. Or, use an LLM to generate an image and run that through Metasynth. That would at least be taking the AI content out of the medium in which it is derivative and into one where it might actually spark some new insights.
As with all AI related discussions I think it’s worth mentioning that judging and writing off the potential of a technical application based on the currently existing tools is probably a bad idea. LLMs are low hanging fruit that has become commercially successful thanks to the availability of large amounts of data and processing, it’s kind of a brute force method. The shortcomings are widely known and other kinds of smarter and more efficient systems will inevitably be developed over time.
So I think it’s entirely likely that we’ll have something fast and intuitive, and probably sooner than you might expect. Current AI stuff in general does seem to be a quite clunky and takes a bit of work to figure out how best to prompt it though.
Yes, it could.
I guess I’m of the mind where if you’re sampling something you already have an idea in mind.
I’m totally down with letting AI create stems for someone to get a better sample, but even then you have an idea already in mind and are using AI to better realize YOUR idea.
Ah found it.
took the words right out of my mouth.
Of course, all computers 30 years ago were quite clunky and look where we are now with little flat computers in our pockets so I don’t disagree.
exactly… or online