This is the repository for the paper
Modeling harmonic progressions in symbolic music notation. It encodes and produces musically coherent and varied chord progressions. We employ a transformer-based architecture trained on a comprehensive dataset of 48,072 songs, which includes an augmented set of 4,300 original pieces from the iReal Pro application transposed across all chromatic keys. We introduce a novel tokenization and voicing encoding strategy designed to enhance the musicality of the generated chord progressions.
The dataset is formed with original 4,300 songs from iReal Pro. It contains chords and structural data from musicXML files and metadata (style, time signature, tonality, composer, title). A custom song parser included which extracts bars (measures), chords (root notes, qualities, extensions, slash chords, duration), and form (repeat bars, sectional jumps, codas). Distribution of Styles in the dataset:
Custom Data Augmentation was implemented to avoid music21 errors with enharmonics.
Visualizing the process of song transposition. In this case, we transpose Eb major to C major. The example is a fragment of Giant Steps. The time steps of each chord transposed are shown on a blue scale, from light blue to dark blue. A tuple distance coordinate is used to transpose; the first one is the "line of fifth" (LOF) distance, and the second is the distance in the chromatic scale (CD). The LOF is useful when checking tonality relations, e.g., if the distance of a dominant chord is 13 in the LOF, the chord is corrected to its enharmonic distance -1. It is shown in the Abmaj7, corrected after Eb7 G#maj7
To hear some samples from the voicing model, go to this link
This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0). For more details, see the full license at Creative Commons CC BY-NC 4.0. Or read the license document attached.
If you use this work in your research, please cite it as follows:
@inproceedings{dalmazzo2024chromaflow,
author = {David Dalmazzo and Ken Déguernel and Bob L. T. Sturm},
title = {ChromaFlow: Modeling And Generating Harmonic Progressions With a Transformer And Voicing Encoding},
booktitle = {15th International Workshop on Machine Learning and Music (MML 2024)},
year = {2024},
month = {September},
address = {Vilnius, Lithuania},
url = {https://github.com/Dazzid/chromaflow_1.1},
}