preset-gen-vae

Introduction

This repository provides models and data to learn how to program a Dexed FM synthesizer (DX7 software clone) from an input sound. Models based on Variational Autoencoders (VAE) and results are described in the DAFx 2021 paper and the companion website.

Neural networks and training procedures can be configured using the config.py file.

Project structure

Data

• ./synth/ Dexed presets SQLite main database (> 30k presets), Python modules, and pre-rendered .pickle/.txt/.wav files
• ./data/ PyTorch dataset and pre-computed spectrogram stats

Saved models and logs

• ./saved/runs/model_name/run_name contains Tensorboard data for all models and learning runs
• ./saved/model_name/run_name/config.json stores the full config (model, hyperparams, etc...) of a given run
• ./saved/model_name/run_name/models/ stores trained models (sorted by epoch)

Citation

If you use this work in your research, please cite:

@inproceedings{levaillant2021vaesynthprog,
	title        = {Improving Synthesizer Programming from Variational Autoencoders Latent Space},
	author       = {Le Vaillant, Gwendal and Dutoit, Thierry and Dekeyser, Sébastien},
	year         = 2021,
	month        = Sep,
	booktitle    = {Proceedings of the 24th International Conference on Digital Audio Effects (DAFx20in21)},
	location     = {Vienna, Austria}
}