AW: [sdiy] Vocoder & neural network...

[Nils Pipenbrinck]

Hi Jbv,

I worked a lot with NN's before. I did a lot of OCR and other optical
recognition stuff. I used neuronal networks  for quite a time, but in the
end I always came up with a better solution.

NN look like magic if you first read about or play around them. Once you
demystify their inner workings they're just a very large but also very
simple mathematical hack. Everything you can do with a neuronal network can
most probably better done with some math. Neurnal networks don't think. They
can only be trained to calculate a mathematical equation you are to lazy to
calculate yourself.

The learning feature is imho one of the few reasons I would ever choose to
use a NN again. It's nice if the problem slowly changes, and the NN is able
to change with it.


Using neurnal networks with a only a vocoder won't give you to much
posibilities.

If you tune your filters to the formants of human speech you could put a NN
behind the outputs and train it to recognize the vocals. That's quite easy
to do, but unfortunately all a NN would be able to. The problem is, that a
vocoder input filterbank just does not give you enough information to do
anything better. So you could light up some LED's showing the input vocal,
or use them as controllers or whatever.

However, if you feed a large NN with FFT output of a sampled signal you can
do a more complex things. Pitch recognition is one possible use, but there
are cheaper and more reliable ways to do the same.

Timbre recognition might be an interesting field: Train the neuronal network
to recognize the timbre produced by your different VCO waveforms and
different VCF settings. Once you've done that you are able to feed any
signal into your FFT, and the NN could suggest you a VCO waveform and VCF
setting (or output a CV) that would match the timbre somewhat.

Btw, afaik the Axxon guitar to midi device claims to use neuronal network
technology. Smells like marketing hooey for me. OTOH the machine itself is
nice.

  Nils


> 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...


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

Has this been tried already ?

JB