Hi,

If I figure out how to implement a neural network opcode and how to use it I
have two ideas on how it might work.  One would be to have the inputs and
outputs audio streams.  The other would be to use the spectral variables,
wsig for input and output.  I think the opcode may look like one of the
following:

asig0 [, asig1, asig2...] neuralnet imode, "weights.ann", asig0, asig1, ...

wsig neuralnet imode, "weights.ann", wsigin


The problem with the first one is that I'm not sure if an arbitrary number
of audio signals can be specified.  The second one is probably better.  The
file "weights.ann" would be a text file with a list of the number of
neurons, weights and thresholds something like:

3
4,4,4
w00,w01,w02,w03,th0,w10,w11,w12,w13,th1,w21,w22,w23,th2,w30,w31,w32,w33,th3
w00,w01,w02,w03,th0,w10,w11,w12,w13,th1,w21,w22,w23,th2,w30,w31,w32,w33,th3

A separate training opcode could be be something like:

     neuraltrain imode, "weights.ann", wsigin, wtrain

Where wtrain is a training signal of the desired output.  This would update
the file "weights.ann" when finished.

Bye,
Hans Mikelson