Datasets

Musical performance datasets for expressive variation modeling, pattern recognition, and groove analysis.

DoMP

MIDI Monophonic Melodic
4,392 files 40 performers 333 patterns Piano & Guitar

Dataset of Monophonic Patterns. MIDI recordings of monophonic melodic patterns performed by 40 anonymous musicians (20 pianists, 20 guitarists). Each pattern family consists of one ground-truth reference and multiple expressive variations differing in timing, dynamics, articulation, and ornamentation. Mean 13.2 versions per pattern (range 4–32). Designed for expressive music generation, variation modeling, and monophonic performance analysis.

DoPP

Audio Polyphonic WAV
2,276 files 20 performers 176 patterns Piano & Guitar

Dataset of Polyphonic Patterns. Mono audio recordings (WAV, 48 kHz, 16-bit PCM) of expressive musical performances by 20 anonymous musicians (10 pianists, 10 guitarists). Each pattern family pairs a ground-truth recording with expressive variations spanning ornamental, harmonic, and structural divergence. 89.3% of variation pairs remain harmonically close (chroma similarity >0.9), reflecting predominantly ornamental variation. Mean duration 5.2 s per file. Designed for audio-domain expressive performance analysis and generation.

DoDP

MIDI Drums Rhythmic
994 files 8 performers 77 patterns Percussion

Dataset of Drum Patterns, v1. MIDI recordings of rhythmic drum patterns by 8 anonymous percussionists using GM standard drum notes (kick, snare, hi-hat, toms, cymbals). Each pattern family pairs one ground-truth with expressive variations differing in timing, velocity, density, and instrumentation choices. Mean 12.9 versions per pattern (range 10–51). Note: files mix ticks-per-beat (480 or 960) and drum channel (0 or 9) due to differing DAW export settings — both are documented in metadata. Designed for drum pattern generation, groove analysis, and rhythmic variation modeling.

DoDP2

MIDI Drums Rhythmic
2,177 files 10 performers 100 patterns Percussion

Dataset of Drum Patterns, v2. Complementory to DoDP, featuring 10 anonymous percussionists, uniform MIDI encoding (96 ticks/beat, channel 0), and compact loop patterns (mean 3.9 beats, max 6.9 beats). Substantially denser sampling: mean 21.8 versions per pattern (range 19–59). Introduces artist-level template patterns (patternID = -1) — 194 files representing each performer's baseline rhythmic character, usable as style anchors or zero-shot conditions. Designed for groove modeling, style transfer, and densely-sampled rhythmic variation research.

All datasets share a unified artist ID namespace and flat file naming scheme (artistID_patternID_versionID). For related publications, see the Publications and Research pages.