Dilemmadata: A symbolic dataset for music research

This repository contains the processed and aligned data from two major annotated music corpora: the AugmentedNet dataset and the Distant Listening Corpus (DLC). The data has been preprocessed and converted into pitch arrays — tabular representations suitable for graph-based machine learning models used in automated music analysis tasks.

Overview

This project serves as the data infrastructure for training graph neural networks (GNNs) on multiple music analysis tasks, including:

Cadence detection (identifying cadence types in musical passages)

Phrase segmentation (marking phrase boundaries)

Key analysis (local and global key detection)

Harmonic analysis (chord quality, inversion, root, bass note)

Roman numeral analysis (functional harmonic analysis)

Rhythmic analysis (downbeat and metrical analysis)

Voice leading (analysis of voice leading patterns)

Section segmentation (identifying structural sections)

Pedal point detection (sustained bass notes)

Note degree inference (scale degrees relative to local key)

This resource has been demonstrated through the AnalysisGNN framework [Code][Paper] and serves as a foundation for training neural networks on automated music analysis tasks using multi-task learning and graph-based representations.

Data Sources

1. AugmentedNet Dataset

Source: github.com/napulen/AugmentedNet

AugmentedNet is an automatic Roman numeral analysis neural network developed by Néstor Nápoles López as part of his PhD research. The dataset includes:

353 pieces from multiple collections (Beethoven Piano Sonatas, Bach chorales, TAVERN, etc.)

Roman numeral annotations for harmonic analysis

MusicXML scores with RomanText annotations

Split: Pre-defined test/training/validation splits (v1.0.0 dataset)

Key features:

Cadence annotations (cadential labels)

Roman numeral analysis (functional harmony)

Chord annotations with inversions

Synthetic training examples via texturization

Reference:

Nápoles López, N., Gotham, M., & Fujinaga, I. (2021). AugmentedNet: A Roman Numeral Analysis Network with Synthetic Training Examples and Additional Tonal Tasks. In Proceedings of the 22nd International Society for Music Information Retrieval Conference (pp. 404–411). https://doi.org/10.5281/zenodo.5624533

2. Distant Listening Corpus (DLC)

Source: github.com/DCMLab/distant_listening_corpus

The Distant Listening Corpus is a large-scale collection of annotated musical scores from the DCML (Digital and Cognitive Musicology Lab) corpus initiative. It includes over 40 subcorpora spanning music from the 17th to 20th centuries:

Bach, Beethoven, Chopin, Mozart, Schubert, etc.

Comprehensive harmonic annotations using the DCML standard

MuseScore 3.6.2 files with embedded annotations

TSV exports of notes, measures, chords, and harmony labels

Included subcorpora (selected):

beethoven_piano_sonatas, chopin_mazurkas, mozart_piano_sonatas

bach_en_fr_suites, bach_solo, schubert_winterreise

debussy_suite_bergamasque, grieg_lyric_pieces, liszt_pelerinage

monteverdi_madrigals, scarlatti_sonatas, wagner_overtures

And many more...

Key features:

Phrase boundaries

Cadence annotations

Local and global key annotations

Pedal point annotations

Section start markers

Note degree annotations (scale degree relative to local key)

Reference:

Hentschel, J., Rammos, Y., Neuwirth, M., & Rohrmeier, M. (2025). A corpus and a modular infrastructure for the empirical study of (an)notated music. Scientific Data, 12(1), 685. https://doi.org/10.1038/s41597-025-04976-z