Machine learning audio effects

Trying to learn machine learning for audio effects. This is my own take at coding audio effects using machine learning. In OCaml.

It started with the LSTM-based approach described in

• this github repository
• this article
• this blog post

Usage

In order to learn an effect you need both an original sound file (say source.wav) and a processed sound file (say target.wav). You can start learning with

leffect -s source.wav -t target.wav

This will learn the effect. While doing so it will also produce an output.wav file (name can be changed with -o) resulting of the processing of the input file by the current network. Some other useful parameters are

• --rate: the learning rate
• --size: the size of the network
• --play: play the output in realtime

At the end, it produces an effect.json file (the name can be changed with the --json flag).

In order to process a file with an already learned effect (whose parameters are stored in effect.json), you can simply do

leffect -i file.wav -o output.wav