NotaGen: Symbolic Music Generation
electricalexis.github.io
electricalexis.github.io
It’s just like LLM generated prose to me. The orchestration and motif development is musical sounding, but there’s these artifacts that are like the “It’s important to consider” or “Would you like further suggestions” of the aural world.
I think it would be fun to write up a detailed analysis of a full piece.
The waltz sounds like elevator Chopin with all the bittersweet irony sanded off (and couple of questionable harmonies.)
It's an achievement to get this far, but it's still very uncanny valley, and (IMO) a fundamental limitation of LLMs that can't see beyond the superficialities of weightings into what the structures mean.
Suno et al are moving in this direction but I honestly think development will be somewhat stunted until we get a good open source model(s), and something like control-nets.
Classical music isn't special superior form of art, and people who try to paint it as such tend to be elitists who want to have something to hold over other people to make them feel superior. Kind of like how people latch on to wine to try and seem sophisticated (ironically, wine snobbery has its origin with kings and chieftains of conquered territories trying to fit in with their Roman oppressors, but at least they were cognizant of why they were doing it, unlike modern wine snobs). There's plenty of metal that's far more technical and complex than typical classical music, but you don't see metalheads going around telling people to listen to "real" music.
This means that classical music, more than other traditions, is a natural target for both broader academic study as well as automated generation by AI's trained on some sort of existing repertoire.
I just mean to say, I don't think there is a clear divide between a classical mentality for composition and a modern mentality. It is just we study classical hundreds of years after the fact and that academization of the music has lent a particular view to it. If we study metal in a hundred years the same way we study classical (or jazz) it may seem just as rigid.
Chopin predates audio recording, so it’s not as if someone recorded his performance and then transcribed it like people do today with jazz solos. I’m assuming that Chopin was the one to transcribe and publish after he had a good idea because few would have the recall to do so accurately.
Sometimes compositions are generated spontaneously by the composer (one pass) and sometimes they require extensive labor and refinement.
It’s not to say that classical music doesn’t also include improvisation, and I agree that there’s not a clear divide between classical mentality and modern sensibilties. Although today, the composition (if not scored) ends up taking it’s final form as an audio recording. More and more the composition exists untranscribed in the DAW session.
The argument against my point would be, but the people learning the music (i.e. studying it) later started with the written music and went from there. And that is a differentiator for their own development as composers. That would be compared to modern musicians studying modern pop music where they start with a recording and go from there.
My point is that the differentiation is in the contemporary study of music and less in the method of composition.
As I explained in another response, I see the literate nature of classical music as related more to the transmission of knowledge from master to novice and much less in the process of composition. In the past, music was written down for study. Today music is recorded and studied directly. I have no doubt this has consequences on the student, but I don't think it has as much consequence on the process of composition itself.
The Barry Harris approach to jazz BTW is very much informed by what we know about the way classical improvisation worked, so even though it's transmitted by video there is in fact a link to the scholarly tradition. (The style is of course different, so these aren't quite the same thing! But not that far either.)
I'm arguing that reasoning "classical music is important because it is written" is backwards. It is written and studied because people think it is important. The degree that the formalisms are applied is directly related to how important the elites in the academies think the music is. Please don't mistake that claim with me suggesting pop music or metal are in fact as important as classical or jazz. I'm just pointing out that nascent formalisms for those genres obviously exist.
But my main point is, just because classical music has been deemed important enough by the academy to write it down, study it and enforce its strict reproduction, that does not imply that it is easier for an AI to learn the genre or reproduce it effectively. I have no doubt that just like millions of humans, multi-modal AIs will be able to use the vast library of recorded music in all genres to easily reproduce those genres compositions perfectly well. There is no privilege to classical just because the historical context meant it was committed to paper rather than record.
This is of course silly; there are important traditions of broadly non-literate music, and modern popular music (with its huge variety of "genres", including metal, EDM etc.) clearly qualifies. I have only argued here that classical music being written makes it special/unusual, in a way that's legitimately compelling to some. (Including academic elites, and people looking for stuff to train an AI on.)
Do note that classical music being "written down" is not something that has happened "after the fact": the written form is how the pieces are published to begin with! (There's an interesting contrast here with 'folk' tunes, that spread orally in many subtle variants and are only written down afterwards.) Now, it is also true that performance practice can add a lot, historically; you don't have to reproduce strictly what's written. But the abstract "blueprint" to what you're performing is given by the written piece.
> I have no doubt that just like millions of humans, multi-modal AIs will be able to use the vast library of recorded music
Training an AI on recorded music is really really hard. They have to learn about how the acoustics of every single instrument works and this overwhelms the information that we actually care about, of how a piece of music goes. The difference there is absolutely clear, and quite massive.
> you don't see metalheads going around telling people to listen to "real" music.
Have you _met_ a metalhead? :) saying that as one myself (less now, used to play in bands and stuff), this is one of the snobbiest group music-wise!
Like when we were 15, yeah sure, everything sucked, Opeth rules, Mayhem is tru Black Metal! etc. Though later on only my really immature friends kept those opinions. But I think this goes for any niche subculture that people use to define themselves. I too am an old metalhead, still practicing! Dress like a 15 year old, crank the volume and bang your fucking head!!! \m/(>.<)\m/ ;-)
I am generally in favour of throwing shade on LLM "creativity", at least thus far, but this is an uncomfortably strained flex. Literally.
We need an AI DAW, not Suno.
As long as there’s music made by real people this AI music stuff will only be interesting to the very uninspired, the people that will hype up anything that’s AI and commercial projects who don’t have to pay royalties for the music in their advertisements, elevator, phone waiting music and whatever else they need a quick tune for.
Don’t get me wrong, the technology is nice but I don’t see myself consciously enjoying AI music, at least not till I start dementing or smth.
The major issue I can foresee going forward is that the only public domain sheet music available is in the classical idiom. And people don't really care about writing new pieces in a Baroque style. It's a fun exercise, but not really relevant to modern music. The "pop music" listed in the article sounds like "Baroque-pop" presumably because that's what is in the training data. That's why a DAW style system would be more useful and welcome, although there again - what data do you train on? There aren't just millions of publicly available "EDM DAW files" laying around to train on.
I'm not sure that this is accurate. You can take a look at IMSLP and see how much music there is that's in a more "popular" style (while still not anywhere close to modern pop, of course). Especially vocal music (most commonly songs) and instrumental dance pieces, which were the most common choices among 19th century amateur performers playing music in their home salons and parlors. It takes a lot of work to accurately transcribe the music from score notation into a form that AI can be trained on, but that work too is happening via e.g. the "OpenScore" effort. So I think we can expect better AI-generated notated music in the future. It could become a nice "toy" domain for sequence modeling (including language modeling) architectures, where high-quality results might be achievable with very little training effort. A bit like MINST is today for image recognition.
My own opinion, as an amateur musician with no classical training, is the music I heard on this AI generated page would not have stood out to me compared to the rest of the pieces on that channel. I would wager that in a randomized test of these average middle of the road composers vs. an AI score generated by NoteGen that almost no average person would be able to distinguish them.
That leaves two categories. The highly trained individuals (e.g. professional composers or well educated amateurs) who may be able to tell a difference (I would love a real study on this and not just "vibes"). And the master composers (Beethoven, Chopin, Brahms, Liszt, etc.) which stand above the average human composers of their own times.
I would admit the pieces composed by the AI aren't at the level of the master, but they do feel indistinguishable to me from the average.
None of the pieces I heard were good, but I think they'd get a passing grade in a composition class.