Re: [Om-synth] Re: Om problem, today's CVS

[Ross Clement]

On Sat, 2006-02-25 at 22:11 -0500, Dave Robillard wrote:
> On Sat, 2006-25-02 at 07:48 -0500, Dave Phillips wrote:
> > Yes, it 
> > needs more patches. I'd like to help out and design some, but my time is 
> > socked. I had hoped to make some patches for Sean Bolton's WhySynth too 
> > but I just don't have the time. I suggested to Sean that he consider 
> > building a patch randomization utility into WhySynth. Would something 
> > similar be possible in Om ? Seems like it kind of goes against the grain 
> > of its design, but it doesn't hurt to ask. :)
> 
> Statistically speaking, the chances of generating a patch that actually
> produces noise (let alone pleasant noise) are pretty much zero :)
> 
> If it's completely random anyway.. if it were much more clever, you
> never know.  Playing around with stuff like this should be pretty easy
> with the Python or SuperCollider bindings at least...

My actual speciality is Artificial Intelligence. Making an automatic
patch/sound generator for synthesisers is something I've always thought
of doing. But it would be just far, far, too big a job to do even
halfway reasonably for a random patch generator is a big, big, job. And
to have a system (as I'd like to do in my pipe dreams) that could take a
natural language ("English") description of a sound and synthesise it is
even bigger. I think I could produce a design for such a system, but the
amount of real-world knowledge that would need to be encoded in such a
system to make it work is just too large. I believe that it would be a
major international effort to get a good AI patch generator that could
respond to descriptions of sounds. Currently many people in many
subfields of AI are trying to ignore the central difficulty of AI by
concentrating on learning systems, frequently IMHO quite weak and simple
learning methods such as Neural Networks. But the amount of training
material that would have to be created to train even much more
sophisticated learning methods would be just too large a task in itself.

Something Dave says matches something I've wanted to do for a while. He
mentions the problem of a random patch not even creating a sound. Unlike
making "good" sounds or sounds that match a description, it's possible
for a program to detect by itself whether a patch makes any sound at
all. Hence it would be possible to create a machine learning program
that learns rules predicting when a patch will make no sound. E.g. if a
lowpass filter with cutoff X is followed by a highpass filter with
cutoff Y: Y>X then no sound. But in reality envelopes and other
complexities will make the design of language for describing sounds
again very difficult. Just using the straight parameters and module
names is probably not going to work, as the reasoning will have to be at
some higher abstract level. Not that I can think of what that level
should be.

I do have some simpler experiments planned that I'd like to do "some
day", but so far that day has never come up.

Sorry to all if I'm off-topic here.

Cheers,

Ross-c