Music listens
The project, in brief, is to create a program that creates highly differentiated large-scale musical works. This involved & involves the development of a full-blown computational theory of music. I have a nightmarishly huge pile of undocumented material which I would like to expose to the world. But to do this I need first of all to establish the artistic & intellectual background of Jack & Jill
So allow me to present to you some of the issues involved in autonomous music creation. The first thing to notice is that the problem is positively correlated with your musical standards. If you’re ok with listening to hours of white noise, then if you’re on linux, the problem has already been solved for you:
cat /dev/urandom > /dev/dsp
and you may next perhaps wish to build an aphasic chatbot. Perhaps some people, when they think of computer-generated music, imagine a computer making works in the “style” of classical composers like Beethoven. There’s a microsoft generative music program floating around that offers a ‘classical’ style which seems to always generate something like a jumbled version of Für Elise, which, I guess, would meet some people’s expectations. But is it possible for the computer to invent an 11th symphony of Beethoven at the relevant high technical & expressive level? Rather tough, because Beethoven was an exploratory composer whose major works do not tend to repeat previous things: each work can be a world all by itself. The Waldstein sonata has a completely different feel from Op. 101, or 110, or 111, or any of the other main Beethoven works. And the difference applies to all other works by all other composers, too. So to make a piece by computer-Beethoven you must ensure that it’s different from everything else. As a matter of fact, the problem of uniqueness is a vital component of musical AI: the composer must know what it’s done, what would constitute a difference in something from everything that has been done, what would constitute a trivial variation – the Appasionata with 3 notes changed, probably for the worse — or a systematic variation (repeat all notes twice) which brings nothing positively new, etc. I’ll return t this again and again, because, as I’ll eventually try to show, the relation between the existing and the new is a precept that digs into the technical fiber of music, at least as I theorize it.
For this reason I regard the “style copy” problem of AI as the kind of problem that doesn’t pose the right questions: the idea of style seems to be about some sort of constancy that one would feign think of as reusable components. It doesn’t begin to address the general problem of difference, nor of how components are supposed to fit together, in other words what the logic of a piece of music of a piece of music is supposed to be like, and how that logic plays with the musically other.
And yet the style copy issue does raise one point that to my mind is uncircumnavigable: the technical challenges in the qualitative distinctions between older ‘classical music’ & newer ‘modern music’. I was just at a talk by music psychologist Lola Cuddy, who has been studying the musical memories of people in various stages of Alzheimers. Cuddy observed that musical memories were nearly intact when all else had left, including the power to detect and wince at wrong notes. The music she played was of the old standard type, the Wagner wedding march & that sort of thing. I did not bother to ask her why she did not investigate whether any of her subjects could recognize electronic compositions by Stockhausen. My point is simple: older music has cognitively interesting properties which ‘modern’ music doesn’t tend to have. And in trying to get a computer to think about music, I think it is crucially important to try to capture some of these ‘older’ properties, including the sense, for example, that what you are listening to has very good melody. In trying to make musical AI happen, it is necessary to aim for a level of universal accessibility: the AI succeeds when people, including sophisticates & non-specialists, are able to enjoy the music on their own terms. In other words, AI cannot be AI unless it takes the world as it is found. I do not take music as found – because of the role of the new in all music, even the most convention-ridden. What I do take as found is the issue of listening potentiality.
I’ll address this sideways. If you’re interested in computer music then you know all about algorithmic music, in which the idea is just to get a program to generate something with properties that you like. Now the crucial thing here is that the computer has no idea what it’s doing – that is, there is no internal theorization of what the music it is grinding out sounds like. And that is bad.
It is bad in virtue of my old aphorisms: music listens. What this means is that within music there are responses as if the music was internally aware of how it sounds & feels. It feels bad, but another part says “there there,” and another says, “hope is forlorn!” — in Mahler or something. The one thing knows about the affective feel, the potential meanings, of the other thing. You are free to hear music as if the music were engaged in a conversation with itself, between its parts, between its parts and the whole, etc. That is how I heard Bach’s music as a kid: it seemed like a conversation undertaken in a super-intelligent mode, & that is because the parts are listening to and responding to one another. The average listener probably does not hear music in this way – indeed I assume we all hear music in our own ways – but I believe that any level of musical intelligibility is dependent on some sort of integration of the music-internal discourse. Your favorite singer hits a particular note in relation to the way things were such that this note is about moving forward into that particular response: it feels right to everyone. It feels right because the music has listened to itself and has expressed the way in which it has done so. At least that is the operational listening hypothesis I am presenting.
To grasp, or model, listening potentiality, I must theorize at least the way in which the music could be said to have been listening to itself – even if this is somewhat hypothetical.
Filed under: Uncategorized
