Resolved yes if it beats human music at least 50% of the time in a general audience or expert poll across at least 10 samples of different prompts.
(It might seem weird for me to be the one to ask since I have one of the largest NO positions, but I'd rather ask this myself than have someone predicting YES bring it up, which probably would have happened at some point.)
I'm starting out with an initial NO bet WAG not from having reviewed the literature or having read the comments yet, but from just my initial instinct that music has so many complicating variations across different genres, that it will have a sort of tinnyness to it which will take a while to resolve. However, if a research team comes up with a particular benchmark aimed at a single genre, then I could see this resolving within 2023.
So...will this market accept domain-specific or genre-specific music generators or does it have to include for example, 100 different genres? Does it need to be a generalized music maker? Obviously a generalized music maker is much more impressive, but harder to build.
Another question - does the AI generator constitute merely a large language model similar to how GPT text and image transformers operate? E.g. is it a zero-shot music generator? Or, since you mention, "AI" rather than something specific, e.g. a GPT, then can the AI music generator include some kind of, "listening filter," on the output side which eliminates all clearly horrible or nonsensical sounding music? What if that listening filter includes human intervention / mechanical turk style cherry picking...? Is it still an AI?
ChatGPT includes a lot of cherry picking and human intervention, in the form of constant fine tuning...yet it is still called AI, even though it's human-computer-hybrid intelligence.
These new models will fail on prompts such as "Jazz song with a complex drumbeat where the snare is only hit once throughout the chorus", similarly to how text-to-image models fail when asked to draw some specific number of objects.
Like text-to-image, scale might just be all you need to fix this, but doubt that will happen by the end of 2023.
Do the different prompts need to specify different genres? It seems like a prompt like "Repetitive house music" is much easier to do realistically than genres including singing or more varied compositions
Depends if singing is included
Published 7 hours ago: https://google-research.github.io/seanet/musiclm/examples/
I think the biggest issue now is sound quality, if it was just crisper many of these would have me fooled. Larger dataset + higher resolution on the audio and this resolves YES.
@EricJang I think the generator is what should be available, not the sound? So regardless if it's 30 sec or 3 min clips, what matters is if users can generate new ones?
What I meant was more an indication of advances and work being done in the field. I think the "Bella ciao" samples were especially impressive!
@HenriThunberg This is impressive and promising, but questions remain. How cherry-picked are the examples? Can longer extracts follow a sensible form and structure? Can it stick closer to the descriptions than in the examples? When will a product like this be publicly available?
Idk, something like "For a publicly available model: Given the best possible prompts on ten different instrumental genres, the hit rate is less than xx% in guessing whether it's AI output or not when paired with human-created samples, from a general audience (say Twitter or Manifold poll)."
This seems like quite a high standard; I don't think we've achieved this for other nontrivial generative tasks.
@osmarks I mean we’ve had ai music generation capable of being indistinguishable from human generated music for years now haven’t we, I don’t think adding a requirement of text based prompts is that high of a standard, assuming someone actually tries I think music generation will be one of the easiest nontrivial tasks to reach that standard.
@enemel i'm was specifically thinking of Emily Howell which has been around since at least 2009, i'm seeing that actually theres some debate about whether it has passed the Turing test, and many of the original sources cited no longer exist, but it's good enough that I believe it could have passed a Turing test.
My skepticism here is mostly just that I think it's reasonably likely that no major lab will try in the next year. On a technical level I think this is very doable.
Are we talking about instrumentals only here, or realistic sounding voice+lyrics (which sounds much harder)?
@Nostradamnedus Also what does "most of the time cannot be told apart from human musicians" mean - at the low end we might have "most people listening to it out of context, especially a small snippet, would not immediately suspect it's AI", and at the high end we have something like "music experts given one obscure human piece of music and one AI one can't identify the AI one much better than 50% chance"
Will a text prompt based AI music generator that most of the time cannot be told apart from human musicians be publicly available by the end of 2023?, 8k, beautiful, illustration, trending on art station, picture of the day, epic composition