Re: robot com-posers


Subject: Re: robot com-posers
From: Eliot Handelman (eliot@generation.net)
Date: Wed Feb 16 2005 - 23:33:13 EST


>> What I what is something that actually knows
>> about music, not a magic number
>> filter that explains nothing.
>>
>> -- eliot

> I have an idea for you. Theres alot of research being done involving
> engineers trying to get computers to speak.

To understand natural language, I think you mean.

> There are many languages as there are many types of music. They are
> both aural or written forms of expression. They both, to a great
> degree express a cultural background. They both involve the
> understanding of sound by recognising patterns, and combinations of
> patterns...

Ok, look. I ran out of time, and didn't get a chance to bring together
various threads I'd set up, in particular, I didn't have any time to
talk about rhythm, its role in the old/new dynamic, or in ways
patterns can be construed as rhythms, or what they're patterns of, and
so on. I mentioned some fairly complex patterns of transformation that
occur in Happy Birthday (eg the "parallel fork") and how their
transformations can be reinterpreted as an affective narrative (recall
that I suggested a kind story-telling in all this). but I didn't have
time to discuss the problems involving how to type patterns,
generalizing them, etc. I didn't have time to discuss actual programs
that I've written that explore different approaches to getting all of
this together. I didn't have time to show in what way listening is
always active, in fact, compositional -- though I did mention that the
my work on computational listening requires now a composing module
which I'm now working on. I mentioned in previous posts -- since I
wanted to congratulate myself publicly on having finally turned the
problem into something I can almost understand -- that this would be
my fifth shot at the problem. It's not because I'm that stupid -- it's
just a really, really hard problem which no one in the world seems to
have any idea about at all. I regret that I didn't have time to
explain any of this because whether I get it or not I believe this
area is cutting edge and really exciting, and not second to any other
problem that I can wrap my head around or read about in any of the
areas that I'm interested in. I feel a real sense of regret that I'm
not in a position to teach or share what I think I've learned about
music over the years. But please don't try to extrapolate too much
from what I said. For better or worse I;ve been thinking about this
stuff for a long time, and I find it very difficult to condense what I
have to say into a two hour lecture or 500-word abstract.

I left the class with the following idea, which I didn't have the time
to demonstrate: that our musical hearing is massively parallel. This
means, that "patterns" in music aren't necessarily about the kinds of
things that we know how to do pattern recognition with. It's a whole
new area and I'm afraid there's very little work that I've found
useful in this respect.

> MY idea is to focus on a certain type of music, as one would focus
> on one language at a time when teaching a computer how to
> speak. Build a big database of rock and roll and teach the computer
> to recognise key percievable elements of that type of music.

Which are? I spent an hour talking about Happy Birthday. When we ran
out of time the whiiteboard looked like a battlefield and I'd only
just set up the basic elements of motion.

I'm not
> saying its going to be easy. you would have to disect the songs to
> their elements. Teach it to segregate the drums from the guitar and
> so on,

That's a separate area which many people are working on but which is
always going to problematic until the compyter can follow the music
the way you or I do. What I'm trying to do is reated to "following."
So you can't say "first do this." YOu need a model of what's making
sense or adding up in the music.

> Second, teach it to recognize modes progressions. rhythmic patterns,
> timbres... but on the other hand it would have to recognise the big
> picure: tempos, style, overall form,,,the list should be very long.

It's not a list that you're looking for at all. You need to know how
the rhythm affects the pitches. You need in general to know how things
cohere. You will remember that I spoke constantly about details
vs. whole picture and the problem of creating representations in which
small coheres into big.

> Then the computer could design a piece using a huge matrix, each value
> being
> one of a million other values in its category but a problem emerges
> here.

I'm afraid this is a fantasy that won;t work -- I don't believe that
you'll ever get anything but a meaningless jumble which may
accidentally have a moment of coherence. Anyway this doesn't come
towards an explanation of how music works. It's a system for
generating signals that someone has to interpret as music. My interest
is to get a program to do that interpretation. It's also not just to
produce music, but to do that in a certain way.

> you like it?

No. But if you do your thing it's more likely to fit in with the
dominant currents of computational music thinking than my stuff
does.

-- eliot



This archive was generated by hypermail 2b27 : Sat Dec 22 2007 - 01:46:06 EST