README.md

Generating Musical Synthesizer Patches with Machine Learning

This repo accompanies this blog post. The premise is to generate high-quality presets for the Ableton Analog synthesizer in a particular style automatically using generative machine learning models.

Setup

This repo currently requires Python 3.7.9 until Tensorflow can be migrated to version 2+.

Feel free to acquire this interpreter via your preferred method. One way to do so would be

1. Install pyenv
2. pyenv install 3.7.9
3. pyenv global 3.7.9
4. python -m venv venv
5. source venv/bin/activate
6. Pip install -r requirements.txt

Running Training

To run training you can use the run_gan.py script like this:

python run_gan.py --config-file data/configs/config_240_cgan.json

It also supports various CLI options including hyperparameter search. Run python run_gan.py --help

By default, this will output models, plots, and presets to data/generated in the project directory.

Generating Presets Using the Model

To generate presets from a model, use the generate_from_gan.py script. Example usage:

python generate_from_gan.py --config-file data/configs/config_240_cgan.json --model-path data/configs/config_240_cgan_generator_model_e1000.h5 --dest-path . --n-samples 10

To evaluate presets, you will need Ableton. You can get a 90 day free trial here

Running Tests

Test coverage is currently very limited. To run the tests use python -m pytest tests/

Architecture

This project uses the Keras framework (built in to Tensorflow) to construct and train the neural network.

The general flow is as follows:

1. Read training data directly from the repo at data/analog_library (theres not much data available sadly)
2. Parse presets into vectors (logic in presetml/parsing/ableton_analog.py)
3. Run training (logic in generation/gan.py)