zenberry/midihum

MIDI humanisation with machine learning

midihum

midihum (the tool formerly known as rachel) is a command-line tool for humanizing MIDI -- that is, for taking as input MIDI compositions and producing as output those same compositions with new velocity (loudness/dynamics) values for each of the contained notes. It uses gradient boosted trees, with ~400 engineered features, and is trained on 2.6K competition piano performances. For more information, see this blog post.

midihum requires Python 3. It has been tested on macOS Ventura 13.0.1 and Debian 11.

Usage

Using midihum is easy. First clone the repository, navigate to the midihum/ directory, and install dependencies:

pip install -r requirements.txt

Then -- making sure you're still in the midihum/ directory -- simply:

python main.py humanize /path/to/file.mid /path/to/humanized_file.mid

(Or use python3 instead of python if that is your Python 3 binary.)

Performance

The midihum program performs surprisingly well, at least for solo piano works of roughly the type it was trained on, i.e., from the Baroque, Classical, and especially Romantic periods of Western art music. Here are the true velocities and the predicted velocities for nine randomly chosen, not-cherry-picked performances from the validation set:

In the above plot, each dot is a note on event, randomly sampled from the piece. For the same nine pieces, the model also captures the dynamics over the course of the composition well:

Model

The program uses XGBoost gradient boosted trees for its model, where each observation is one MIDI "note on" event (see midihum_model.py), with a large (~400) set of derived features (see midi_to_df_conversion.py).

Acknowledgements

The model is trained on 2,579 performances from the International Piano-e-Competition for pianists aged 35 and under. The midihum tool is dedicated to those talented young performers.