ARMLP

Autoregressive Multi-Layer Perceptron - A neural network UGen for experimental sound synthesis in SuperCollider.

Overview

ARMLP is a small neural network that learns online at audio rate. The network continuously adapts to predict the input signal, and the learning dynamics themselves become part of the sonic output.

Building

Requires SuperCollider source (for headers) and CMake 3.12+.

cd ARMLP
mkdir build && cd build
cmake -DSC_PATH=/path/to/supercollider ..
make
make install

Or use the build script:

export SC_PATH=/path/to/supercollider
./build.sh

macOS architectures

Universal binary (default):

cmake -DSC_PATH=/path/to/supercollider ..

Apple Silicon only:

cmake -DSC_PATH=/path/to/supercollider -DCMAKE_OSX_ARCHITECTURES=arm64 ..

Intel only:

cmake -DSC_PATH=/path/to/supercollider -DCMAKE_OSX_ARCHITECTURES=x86_64 ..

Supernova

cmake -DSC_PATH=/path/to/supercollider -DSUPERNOVA=ON ..

Install locations

• Linux: ~/.local/share/SuperCollider/Extensions/ARMLP
• macOS: ~/Library/Application Support/SuperCollider/Extensions/ARMLP

After installing, recompile the class library in SuperCollider.

Usage

// Basic
{ ARMLP.ar(SinOsc.ar(100) * 0.3) ! 2 }.play;

// With parameters
(
{
    var sig = Saw.ar(80) * 0.25;
    ARMLP.ar(sig,
        size: 12,
        width: 6,
        learnRate1: 0.02,
        learnRate2: 0.03,
        act1: 0.5,
        sineFac: 2.5
    ) ! 2
}.play
)

See ARMLP.help for full documentation.

Parameters

Parameter	Description
input	Audio input
size	Multi-scale buffer length (fixed)
width	Hidden layer neurons (fixed)
epochs	Training iterations per sample
learnRate1	Hidden layer learning rate
learnRate2	Output layer learning rate
ridge	L2 regularization
act1	Hidden activation (0=tanh, 1=sin, >1=softplus)
act2	Output activation
sineFac	Sin frequency multiplier
momentum	Gradient momentum
errorMode	0=L2, 1=L1, 2=Huber
gradNoise	Gradient noise scale
recurrent	Hidden state feedback (not trained)
maxWeight	Weight clipping
reset	Trigger to reset weights

License

GPL-3.0