Latent Wavetables

An autoencoder for wavetables

This project aims to create a latent space of wavetables, such that exploring them is as easy as hovering over a 2D graph. You can generate wavetable sequences from dragging around the graph.

The model is provided at public/latentwt.onnx in ONNX format, and runs in the browser.

Wavetable licences

The wavetables from the web frontend are from WaveEdit Online, which are CC0 - any wavetable generated from this tool is also CC0.

Training

To train the model, you'll need a corpus of wavetable sequences to feed to the model. This project feeds from the WaveEdit Online, but should be flexible enough to be trained on any other wavetable sequences that you may have.

The model.py is a runnable script with a CLI interface that allows you to train and save the model with your own wavetables.

usage: model.py [-h] -w WAVETABLES [-W WIDTH] [-d {cuda,cpu}] [-e EPOCHS] output

positional arguments:
  output                Output ONNX model

options:
  -h, --help            show this help message and exit
  -w WAVETABLES, --wavetables WAVETABLES
                        Path to the folder containing wavetable sequence wav files 
  -W WIDTH, --width WIDTH
                        Wavetable width (default: 256)
  -d {cuda,cpu}, --device {cuda,cpu}
                        Use CPU/GPU to compute the model (default is detected) 
  -e EPOCHS, --epochs EPOCHS
                        Number of training epochs (default: 50)