Neural audio effects

GitHub · Colab

Abstract

This project aims to find creative audio-effects with neural networks. I use the Temporal Convolutional Network (TCN) proposed by Christian Steinmetz and add some modifications to it:
1) Module IntelligentMerge is supposed to provide the facilities to more complex merge of channels;
2) Module WaveLearner allows the model to learn how to change the effect over time.

Trained example-models below are ready for use inside the Neutone VST-plugin

Demo

Experiment 1 (Bass adder + distortion)

model dilation_growth n_layers n_channels receptieve field parameters
Original Christian Steinmetz' model. Download Neutone model 3 10 27 1333 samples (~27.8 ms) ~22 K
train
source
target
examples
Description Original Processed
electric guitar
voice
piano

Experiment 2 (Reverb + distortion)

model dilation_growth n_layers n_channels receptieve field parameters
Original Christian Steinmetz' model. Download Neutone model 7 5 5 33613 samples (~700.3 ms) ~1.7 K
train
source
target
examples
Description Original Processed
electric guitar
voice
piano

Experiment 3 (Reverb + distortion)

model dilation_growth n_layers n_channels receptieve field parameters
Modification of Christian Steinmetz' model with IntelligentMerge module. Download Neutone model 7 5 5 33613 samples (~700.3 ms) ~1.7 K
train
source
target
examples
Description Original Processed
electric guitar
voice
piano

Experiment 4 (Echo)

model dilation_growth n_layers n_channels receptieve field parameters
Modification of Christian Steinmetz' model with IntelligentMerge and WaveLearner modules. Download Neutone model 7 5 5 33613 samples (~700.3 ms) ~1.7 K
train
source
target
examples
Description Original Processed
electric guitar
voice
piano