AI Music As 'Bemusent Park'

 

AI music can't hold a candle to music historically. As interesting as it is for some purposes, such as trying out lyric ideas in different genres, it simply doesn't have the vocabulary that skilled artists have. Sometimes the candle holder will have a brilliant idea, and they can easily produce it mixed and mastered, but when you analyze it, there isn't much substance.

My two main rants about AI music are its narrow harmonic vocabulary and the way lyrics are mapped rhythmically. This mapping consists (as I understand it) of "packing" lyric lines to fit symmetrical schemas that exist in the rhythmic stems, or phrasing in clunky machine-like ways. People do it too--the popular one being the way Lennon sings "people" in Imagine. "People" is naturally spoken on a downbeat. The more natural way is to syncopate it. AI vocals are always doing this, and you can't easily change it.

Discussing the history of songwriting in this way is depressing vis-à-vis how we have understood words and music, evincing two people at the piano working through a song. [The above infographic, generated with AI, is emblematic of AI-generated content in general: looking polished but is essentially "slopped" together from things "on hand" in the LLM].

With AI music, you can't tell the lyricist to sing it a certain way and demonstrate it, save for tweaking the prompt or chatting with it until you get what you want. Imagine Ira [Gershwin] asking George, "is this something you had in mind?" [a nice line in itself], and George would say, "Why are you asking me that? Aren't you the lyricist?"

This whole idea of chatting with bots to get what we want is the strangest thing to me as a creative person. There's a spiritual sloth about it because you're not actually doing anything except placing an order as you would for lunch. This is because we don't have the skills, and AI bridges the gap, but I wonder what's falling through, or if we don't think that's an issue. As to music production, I have neither the gear nor players to produce an album of songs with vocals and AI bridges that gap. But as I take that bridge to the other side, it's not that great a place. It's the park designed with ChatGPT and built out by a billionaire with little artistic substance as we know it. It's a bathetic Utopia: how many times can you go to that amusement park before it becomes Dismaland? 

I'm already feeling this with AI music: you like some of the restaurants on the other side of the bridge, but otherwise it's kind of a no-place, like a Dubai or Shanghai with architectural eyesores everywhere. But it's not enough to simply say you don't like something and not find an interesting way to adapt. With AI music, it's the visuals and art direction, or even a form of fiction writing by developing character artists or bands, as I did with Sacred Geometry, a song I wrote in the late 90s and demoed at home with me doing the vocals. I still like the song, but I'm not the kind of vocalist who can sing every set of lyrics I write. I like my version because my chord changes are more interesting. So the AI music shifts the interest elsewhere, and that's really interesting and as creatively satisfying. 

***

Some AI feature requests

• Interpolate of disparate ideas. Think every band member bringing in ideas and having a process whereby AI would find the similarities between them, similar to what LLMs do with embeddings
• Mapping syllables to note values. The way I've been typically working is to start with rhythms from words. Very often they are isorhythms (same rhythms). I'd like to prompt it to map syllables to note values with rhymes and near-rhymes about _________

[Discuss the history of songwriting and how composers and lyricists collaborated]