"Writing about music is like dancing about architecture."
Have people listen to their favorite songs, then play dashboard drums and hum about it. Record these tagged with the actual song.
You'd need a lot, though. Unless you could somehow fuzz the data to create synthetic pairings, it'd take a lot of people recording themselves sounding awful (which could be fun just by itself now that I think about it).
It's baffling this isn't a thing.
And it's not good songs either. It's bottom of the barrel, a penny a dozen kind of music if you know what I'm saying. They probably do this because it's cheaper than paying rights to play an actual radio.
Now, how the staff survives this kind of psychological torture, I don't know. What I do know is that both staff and the customers would be very glad to listen to anything else, including AI generated music. It's already miles better than some "music" we are subjected to... especially pop music, which is somehow more artificial than actual AI generated music.
There are a lot of places where I'd rather might want actual music, even if AI generated, instead of repeating the same sample over and over because you do not have enough budget to spend on it for something that is at least not completely irritating.
What’s your favourite music? Why?
"Plenty of people have been enjoying this category of software." Really? I ask sincerely. Because it seems to me that this category of software is mostly a curiosity / novelty. Are there actual publishing musicians who have used text-to-music as part of their workflow?
My perception is that no one actually uses these tools for real music creation. Hence my point that they are novelty and not actually impactful.
Besides, even a novelty can evolve to something not imaginable today. The fact that we can go that far with text only at this early stage is a sign of things to come. When creating something, there is a place for expressing your intent with language. Maybe not all of it all the time, but we'll see how things will evolve.
The problem with text-to-music isn't really that text is a poor UX for music description, but also that it's almost definitely a one-shot process. And any sort of turn-taking with the user isn't really co-collaborative, it's really just being forced to accept or reject what the AI has generated, with little feedback or control possible.
That's why I think rethinking the entire UX from a musician centric perspective (rather than: "what's the easiest thing for me to specify as a tool-builder? I know, text to music") is such an overlooked endeavor. I'm much more bullish about things like the anticipatory music transformer, where there is novel contribution on the UX and not just the ML: https://crfm.stanford.edu/2023/06/16/anticipatory-music-tran...
You misunderstand the workflow. I'm not trying to create a text prompt that generates hits. It's a way to get ideas and find interesting stuff rather than doing it manually with a guitar or on the piano. Then you have to walk away for some time to clean your sonic pallette and try again.
As a lifelong musician I am absolutely loving all the conversation around this stuff. Musicians steal everything, all the time. The idea of some genius that sits down and pops out a hit in 30 minutes that comes deep from within their soul and has all the meaning attached to it is a myth. We're trying to make things people can connect with. Yes, some musicians are doing it for the sake of art, but not most, and even with a lot of those, it's just an image they project.
When you hear a song, it has probably gone through half a dozen different versions and was worked on for months. These new AI music generation tools will churn out 100 different ideas from you to borrow from and cut the revisions down quite a bit I think.
Here is a short essay about "The Similarity between architecture and dance": https://www.re-thinkingthefuture.com/rtf-fresh-perspectives/...