[sdiy] Vocoder & neural network...

[Richard Wentk]

At 20:49 11/03/2003 +0000, jbv wrote:
>Hi all,
>
>I'm presently reading some litterature about
>neural networks, and was wondering about
>the possibilty to hook a neural network after
>the analyzer section (BPF bank + rectifiers)
>of a vocoder...
>The idea is to use the network for pattern
>recognition, collect the data and then drive
>a synthesizer section made of a limited bumber
>of BPF used has formants...

Pause for thought: to what end? You won't get something out of a network 
unless you feed it data. So which patterns is it going to be recognising?

>The idea might also work with an FFT ouput...

Congratulations. You've just invented the Hartmann Neuron:

http://www.hartmann-music.com/home/

I suspect they use the pattern recognition as a data compression 
technology. You can characterise the sound of an instrument by 
multisampling for pitch and velocity, and interpolating the FFTs that 
result. (Actually the process works best if you reference the FFT base 
frequency for each note to some subdivision of the fundamental of each 
sample, instead of sampling at some nominal hardware-related sample rate 
like 44.1.)

It sounds great, and the resynthesis processor demands aren't too painful 
these days. But it's still a brute force and ignorance approach, and you 
can end up with monstrous multi-gigabyte sized datasets. If you can reduce 
the dataset to a much simpler model by picking out the most audible 
features you're well ahead of the game.

It's worth downloading the demos for this beast. They sound very nice 
indeed, in a digital but still interesting kind of way. I haven't tried one 
in the silicon yet, but next time I'm in London I'll see if I can check one 
out.

Richard