SOUSA: Synthetic Open Unified Snare Assessment

A large-scale synthetic dataset of all 40 PAS (Percussive Arts Society) drum rudiments for training machine learning models on drumming performance assessment.

Dataset Summary

Metric	Value
Total Samples	99,770
Audio Files	99,770 FLAC (96 GB)
MIDI Files	19,954 (78 MB)
Total Duration	~138 hours
Rudiments	40 (all PAS rudiments)
Player Profiles	100
Skill Tiers	4 (beginner, intermediate, advanced, professional)

Quick Start

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("zkeown/sousa")

# Access a sample
sample = dataset["train"][0]
print(f"Rudiment: {sample['rudiment_slug']}")
print(f"Skill Tier: {sample['skill_tier']}")
print(f"Overall Score: {sample['overall_score']:.1f}")
print(f"Timing Accuracy: {sample['timing_accuracy']:.1f}")

# Stream for memory efficiency
dataset = load_dataset("zkeown/sousa", streaming=True)

Dataset Description

SOUSA generates synthetic drum rudiment performances with:

• Realistic timing/velocity variations modeled from player skill profiles based on motor control research
• Multi-soundfont audio synthesis via FluidSynth (practice pad, marching snare, drum kits)
• Extensive audio augmentation (room acoustics, mic simulation, compression, noise)
• Hierarchical labels at stroke, measure, and exercise levels
• Profile-based train/val/test splits ensuring model generalization to unseen players

Supported Tasks

• Performance Assessment: Predict overall score or specific skill dimensions from audio/MIDI
• Skill Classification: Classify performances into skill tiers
• Rudiment Recognition: Identify which rudiment is being performed
• Timing Analysis: Analyze micro-timing deviations in drumming
• Audio-to-MIDI Transcription: Train models to transcribe drum audio

Data Splits

Split	Samples	Profiles	Distribution
Train	67,806	68	Beginner: 18, Intermediate: 23, Advanced: 22, Professional: 5
Validation	13,000	13	Beginner: 3, Intermediate: 5, Advanced: 4, Professional: 1
Test	18,964	19	Beginner: 5, Intermediate: 6, Advanced: 6, Professional: 2

Splits are profile-stratified: each player profile appears in only one split, ensuring models generalize to unseen players rather than memorizing individual performance styles.

Data Fields

Sample-Level Fields

Field	Type	Description
sample_id	string	Unique identifier
rudiment_slug	string	Rudiment name (e.g., single_paradiddle)
tempo_bpm	int	Performance tempo (60-180 BPM)
skill_tier	string	Player skill level
profile_id	string	Unique player profile ID
soundfont	string	Audio synthesis soundfont
augmentation_preset	string	Audio augmentation applied
duration_sec	float	Sample duration in seconds
num_strokes	int	Total number of strokes
split	string	Data split (train/validation/test)

Exercise-Level Scores (0-100 scale)

Field	Description
overall_score	Composite performance score
timing_accuracy	How close strokes are to the beat grid
timing_consistency	Consistency of timing errors
tempo_stability	Stability of tempo throughout
velocity_control	Consistency of stroke velocities
accent_differentiation	Clarity of accent vs. tap distinction
hand_balance	Balance between left and right hand

Stroke-Level Labels

Each stroke includes:

• onset_sec: Actual onset time
• expected_onset_sec: Expected (quantized) onset time
• timing_error_ms: Deviation from expected timing
• velocity: MIDI velocity (0-127)
• hand: Left (L) or Right (R)
• articulation: tap, accent, diddle, grace, etc.

Measure-Level Labels

Aggregated statistics per measure for tracking performance over time.

Rudiments Covered

All 40 PAS International Drum Rudiments organized by category:

Roll Rudiments (15)

Single Stroke Roll, Single Stroke Four, Single Stroke Seven, Multiple Bounce Roll, Triple Stroke Roll, Double Stroke Open Roll, Five Stroke Roll, Six Stroke Roll, Seven Stroke Roll, Nine Stroke Roll, Ten Stroke Roll, Eleven Stroke Roll, Thirteen Stroke Roll, Fifteen Stroke Roll, Seventeen Stroke Roll

Diddle Rudiments (5)

Single Paradiddle, Double Paradiddle, Triple Paradiddle, Paradiddle-Diddle, Single Paradiddle

Flam Rudiments (12)

Flam, Flam Accent, Flam Tap, Flamacue, Flam Paradiddle, Flam Paradiddle-Diddle, Pataflafla, Swiss Army Triplet, Inverted Flam Tap, Flam Drag, Single Flammed Mill, Lesson 25

Drag Rudiments (8)

Drag, Single Drag Tap, Double Drag Tap, Drag Paradiddle #1, Drag Paradiddle #2, Single Dragadiddle, Single Ratamacue, Double Ratamacue, Triple Ratamacue

Player Skill Modeling

Player profiles model realistic skill correlations validated against motor control research:

Skill Tier	Timing SD	Timing Accuracy	Hand Balance	Velocity CV
Professional	6.4 ms	77.5	90.7	0.08
Advanced	14.4 ms	59.9	89.0	0.14
Intermediate	31.6 ms	33.1	85.5	0.23
Beginner	55.5 ms	7.1	80.7	0.28

Audio Specifications

• Format: FLAC (lossless)
• Sample Rate: 44.1 kHz
• Bit Depth: 24-bit
• Channels: Mono

Soundfonts

• douglasn: Clean kit sound
• fluidr3: FluidR3 GM kit
• generalu: GeneralUser GS
• mtpowerd: MT Power Drums
• marching: Marching snare

Augmentation Presets

• cleanstudio: Dry, close-miked
• practiceroom: Small room reverb
• garage: Medium room, natural acoustics
• gym: Large room, more reverb
• marchingfield: Outdoor/field acoustics

Validation

The dataset has been validated against peer-reviewed motor control research:

Validation	Result
Data Integrity Checks	14/14 passed
Literature Benchmarks	8/8 within expected ranges
Skill Tier Separation	Statistically significant (p < 0.001)

Literature References

• Fujii et al. (2011) - Professional drummer timing: ~10ms SD
• Repp (2005) - Trained musician timing: 15-25ms SD
• Wing & Kristofferson (1973) - Timing response model
• Schmidt & Lee (2011) - Velocity coefficient of variation

Dataset Statistics

Timing Error by Skill Tier

Tier	Mean (ms)	SD (ms)
Professional	0.9	25.1
Advanced	1.7	36.7
Intermediate	1.8	58.4
Beginner	8.1	110.7

Score Distributions

Metric	Mean	Std	Min	Max
Overall Score	-	-	0	100
Timing Accuracy	38.4	30.5	0	90.4
Hand Balance	85.8	9.0	36.4	100
Velocity Control	-	-	0	100

Citation

If you use this dataset, please cite:

@dataset{sousa2026,
  title={SOUSA: Synthetic Open Unified Snare Assessment},
  author={Keown, Zak},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/zkeown/sousa}
}

License

MIT License

Links

• GitHub Repository
• Generation Code