/vaes-for-music

Generating music using variational autoencoders.

Primary LanguageJupyter Notebook

Music creation using VAEs

This is a project work for the 2019 course CS-E4890 Deep Learning in Aalto University. Generated new music can be found in the samples directory, both in midi and mp3 format. If you don't want to download the generated samples, you can listen to some examples also here!

The main notebooks for training the models and generating music are

  • 1-vae.ipynb
  • 2-beta-vae.ipynb (not used in the project)

Data preprocessing is done using codes in the notebooks

  • A_filter.ipynb - filtering the raw midi data
  • B_map_instruments - remapping midi instruments
  • C_splice_tracks - splicing the tracks to create the training examples
  • D_listen_to_data - listening to training examples

Below is an example pianoroll for a generated four-track sequence.

example