Python MIDI track classifier and tonal tension calculation based on spiral array theory

1. Clone the repository

   git clone https://github.com/ruiguo-bio/midi-miner.git

2. Create a virtual environment and activate it, using python 3.8 or higher

   python3 -m venv venv

   source venv/bin/activate

   pip3 install -r requirements.txt

3. Download the random forest model from google drive to the current folder

Google Drive link: https://drive.google.com/drive/folders/1OCGTZnxOenp3K351PWtaaqd8OfFn7XkW?usp=sharing

Please refer to the example notebook file for detailed examples.

1. Total tension calculation. It will output three tension measures for the midi file. The (tension measures)[https://dorienherremans.com/tension] [3] are based on the spiral array theory by [1], which includes cloud diameter, cloud momentum and tensile strain. The default tension calculation window size is one bar. The tension calculation length window could be set by parameter -w.

   -w 1 set the window to 1 beat, -w 2 means 2 beat, and -1 is default for a downbeat (1 bar).
   Example:
   tension_calculation.py -i input_folder -o output_folder -k True
   This will run tension_calculation.py on all the file in the input_folder and output the result in output_folder. -k True means it tries to find key change. Default is not, which is not detecting key change. It will try to find one key change in the song which is usually in pop songs, but not the classical songs.

   The key detection method uses music21 package.

   The vertical step in the spiral array theory can be changed by -v parameter, which should be between sqrt(2/15) and sqrt(0.2). The current implementation set it to 0.4.

   files_result.json records the file key and potential key changing time and bar position. The output of three tension measures are in pickle format.

2. MIDI track separator. Based on random forest classifier, it can find the melody, bass, chord, and drum tracks in the MIDI file and output a new MIDI file with such tracks, including one potential accompaniment track. Use -t to specify the required tracks, e.g -t "melody bass" will omit the files without both a melody and a bass tracks detected. The default is 'melody'.
   Example:
   track_separate.py -i input_folder -o output_folder -t "melody bass drum"
   input_folder contains the original midi files, and output_folder is the destination for the new MIDI file.

   Use -f file_path to select one file.

   Use -c cpu_number to select the number of CPUs for calculation. The default is to use all the CPUs available.

   Use -y True to just output the tracks set by the -t parameter. In default, it will output all the melody, bass, chord, accompaniment and drum tracks. If set -y True, it will only output melody and bass track if -t is set to "melody bass".

[1] E. Chew. Mathematical and computational modeling of tonality. AMC, 10:12, 2014.

[2] C. Raffel. "Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-MIDI Alignment and Matching"

[3] D. Herremans & Chew, E. (2016). Tension ribbons: Quantifying and visualising tonal tension. Proc. of the Second International Conference on Technologies for Music Notation and Representation (TENOR). 2:8-18.

If you use this libary, please cite the following work:

Guo R, Simpson I, Magnusson T, Kiefer C., Herremans D.. 2020. A variational autoencoder for music generation controlled by tonal tension. Joint Conference on AI Music Creativity (CSMC + MuMe).

@inproceedings{guo2020variational,
  title={A variational autoencoder for music generation controlled by tonal tension},
  author={Guo, Rui and Simpson, Ivor and Magnusson, Thor and Kiefer, Chris and Herremans, Dorien},
  booktitle={Joint Conference on AI Music Creativity (CSMC + MuMe)},
  year={2020}
}