dilemmadata

A Unified Dataset for Multi-Task Music Analysis with Graph Neural Networks

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

---

Repository Structure

dilemmadata/
├── README.md                          # This file
├── corpora/                           # Original corpus data (git submodules)
│   ├── AugmentedNet/                  # AugmentedNet raw data
│   └── distant_listening_corpus/      # DLC raw data
├── pitch_arrays/                      # Processed pitch array representations
│   ├── AN/                            # AugmentedNet pitch arrays
│   │   ├── test/                      # Test split
│   │   ├── training/                  # Training split
│   │   ├── validation/                # Validation split
│   │   └── dataset_summary.tsv        # Metadata summary
│   └── DLC/                           # Distant Listening Corpus pitch arrays
│       ├── beethoven_piano_sonatas/   # Organized by subcorpus
│       ├── chopin_mazurkas/
│       ├── mozart_piano_sonatas/
│       └── ...                        # 40+ subcorpora
└── processing/                        # Processing scripts and utilities
    ├── utils.py                       # Core utility functions
    ├── requirements.txt               # Python dependencies
    ├── merged_summary.tsv             # Merged metadata from both corpora
    ├── augnet_summary_v100.tsv        # AugmentedNet v1.0.0 metadata
    ├── dlc_summary.tsv                # DLC metadata
    ├── AN/                            # AugmentedNet processing scripts
    │   ├── create_pitch_arrays.py     # Generate pitch arrays from AN
    │   ├── concat_pitch_arrays.py     # Concatenate all AN arrays
    │   ├── data_overview.py           # Compile AN metadata
    │   ├── test_transformation_equivalence.py  # Validation checks
    │   └── ...
    ├── AN_mscx/                       # Assembled MuseScore files from AN
    │   └── labels/
    └── DLC/                           # DLC processing scripts
        ├── create_pitch_arrays.py     # Generate pitch arrays from DLC
        ├── design_test_split.py       # Design test split alignment
        ├── dlc_pitch_array_specs.csv  # Column specifications
        └── ...

---

Pitch Array Format

A pitch array is a tabular representation of a musical score where each row represents a note, and columns contain features relevant to music analysis tasks. This format bridges symbolic music representations and graph-based machine learning models.

Core Columns (Input Features)

Column	Description	Data Type
onset_div	Proportional integer position (in divisions)	Int64
duration_div	Duration in divisions	Int64
onset_beat	Beat position as a fraction	Fraction/Float
pitch	MIDI pitch number	Int64
tpc	Tonal Pitch Class (fifths: C=0, G=1, F=-1)	Int64
step	Note step (C, D, E, F, G, A, B)	String
alter	Chromatic alteration (#=1, b=-1)	Int64
beat_float	Floating-point beat position	Float64
downbeat	Downbeat position in measure	Int64
is_downbeat	Boolean flag for downbeat	Boolean
ts_beats	Time signature numerator	Int64
ts_beat_type	Time signature denominator	Int64
staff	Staff number	Int64
voice	Voice number	Int64

Label Columns (Task Targets)

From AugmentedNet:

• a_simpleNumeral, a_romanNumeral: Roman numeral analysis
• a_degree1: Chord degree
• a_inversion: Chord inversion
• Cadence information

From Distant Listening Corpus:

• chord: Chord label (DCML standard)
• cadence: Cadence type
• phrase: Phrase annotation
• localkey, globalkey: Key signatures (as tonal pitch classes)
• localkey_is_minor, globalkey_is_minor: Key mode
• pedal: Pedal point annotation
• section_start: Section boundary marker
• note_degree: Scale degree relative to local key

Purpose

The pitch array format enables:

1. Graph construction: Notes become nodes; temporal, harmonic, and hierarchical relationships become edges
2. Multi-task learning: Different columns serve as targets for different analysis tasks
3. Data alignment: A common format for diverse corpora with different annotation standards
4. Efficient processing: TSV format for fast loading and manipulation

---

Data Alignment Process

The two corpora were carefully aligned to create a unified training dataset while maintaining a clean test set:

1. Metadata Compilation

Script: processing/AN/data_overview.py

• Extract metadata from both corpora
• Identify overlapping pieces
• Generate summary files:
  • augnet_summary_v100.tsv — AugmentedNet metadata
  • dlc_summary.tsv — DLC metadata
  • merged_summary.tsv — Combined metadata with overlap information

2. Test Split Design

Script: processing/DLC/design_test_split.py

• Exclusion rule: Any DLC piece that overlaps with the AugmentedNet v1.0.0 test set is excluded from training
• This ensures no data leakage between test and training sets
• Result: 2 pieces excluded from DLC (Chopin Mazurkas BI61-5op07-5, BI77-3op17-3)

3. Validation

Script: processing/AN/test_transformation_equivalence.py

• Verify that pitch array transformations are consistent
• Check data integrity across splits
• Validate column specifications and data types

4. Pitch Array Generation

AugmentedNet:

# Generate pitch arrays with train/test/validation splits
python processing/AN/create_pitch_arrays.py

Distant Listening Corpus:

# Generate pitch arrays for all subcorpora
python processing/DLC/create_pitch_arrays.py

---

Column Specifications

Each pitch array comes with a specification file (CSV or JSON) that describes:

• Column names
• Data types
• Purpose (input feature, training label, metadata, etc.)
• Description of each field

Example:

• processing/DLC/dlc_pitch_array_specs.csv — DLC column specifications
• processing/DLC/dlc_specs_specs.json — Metadata about the specifications

---

Related Projects

AnalysisGNN

Repository: github.com/manoskary/analysisgnn

A comprehensive framework for multi-task music analysis using Graph Neural Networks. Supports:

• HybridGNN, HGT, MetricalGNN architectures
• Continual learning for sequential task acquisition
• Pre-trained models available via Weights & Biases

Reference:

Karystinaios, E., Hentschel, J., Neuwirth, M., & Widmer, G. (2025). AnalysisGNN: A Unified Music Analysis Model with Graph Neural Networks. In International Symposium on Computer Music Multidisciplinary Research (CMMR).

AugmentedNet

Repository: github.com/napulen/AugmentedNet

Neural network for automatic Roman numeral analysis with synthetic data augmentation.

Distant Listening Corpus

Repository: github.com/DCMLab/distant_listening_corpus

A modular infrastructure for the empirical study of annotated music, maintained by the Digital and Cognitive Musicology Lab (DCML).

---

Acknowledgments

This work builds upon:

• The AugmentedNet dataset by Néstor Nápoles López
• The Distant Listening Corpus by the DCML (Lausanne, CH) and 10.5075/EPFL-THESIS-10276