SOUSA: Synthetic Open Unified Snare Assessment

SOUSA is a synthetic dataset of 100,000 drum rudiment performances for machine learning. It covers all 40 PAS (Percussive Arts Society) standard rudiments performed by 100 simulated player profiles across 4 skill tiers, with hierarchical quality labels at the stroke, measure, and exercise level.

Dataset Summary

Attribute	Value
Total samples	100,000
Audio files (FLAC)	100,000
MIDI files	20,000
Rudiments	40 (all PAS standard)
Skill tiers	beginner, intermediate, advanced, professional
Player profiles	100
Tempo range	60-180 BPM
Splits	train (68K) / val (13K) / test (19K)
Audio size	~120 GB
License	MIT

Usage

from datasets import load_dataset

# Full dataset with audio (~96GB download)
ds = load_dataset("zkeown/sousa")

# MIDI + labels only (~2.5GB download)
ds = load_dataset("zkeown/sousa", "midi_only")

# Pure tabular labels (~50MB download)
ds = load_dataset("zkeown/sousa", "labels_only")

# Access a sample
sample = ds["train"][0]
print(sample["rudiment_slug"], sample["skill_tier"], sample["overall_score"])

# Access audio (decoded automatically)
audio_array = sample["audio"]["array"]          # numpy array
sample_rate = sample["audio"]["sampling_rate"]  # 44100

# Access MIDI bytes
midi_bytes = sample["midi"]  # raw .mid file bytes

Dataset Structure

Splits

Samples are split by player profile (not by rudiment or tempo) to prevent data leakage. All performances from a given player appear in exactly one split.

• Train: 68 profiles (68,000 samples)
• Validation: 13 profiles (13,000 samples)
• Test: 19 profiles (19,000 samples)

File Organization

The dataset uses three configurations with different levels of content:

Config	Contents	Size
audio (default)	FLAC audio + MIDI bytes + metadata + scores	~96GB
midi_only	MIDI bytes + metadata + scores	~2.5GB
labels_only	Metadata + scores only	~50MB

Stroke-level and measure-level labels are available as auxiliary files:

import pandas as pd
strokes = pd.read_parquet("hf://datasets/zkeown/sousa/auxiliary/strokes.parquet")
measures = pd.read_parquet("hf://datasets/zkeown/sousa/auxiliary/measures.parquet")

Features

Sample Metadata

Column	Type	Description
sample_id	string	Unique identifier
profile_id	string	Player profile UUID
rudiment_slug	string	Rudiment name (e.g., single_stroke_roll)
tempo_bpm	int	Performance tempo (60-180)
duration_sec	float	Performance duration in seconds
num_cycles	int	Number of rudiment repetitions
skill_tier	string	beginner / intermediate / advanced / professional
skill_tier_binary	string	novice / skilled
dominant_hand	string	right / left
num_strokes	int	Total strokes in the performance
num_measures	int	Number of measures
soundfont	string	Soundfont used for synthesis
augmentation_preset	string	Audio environment preset name
augmentation_group_id	string	Links augmented variants of the same MIDI

Audio Augmentation Parameters

Column	Type	Description
aug_soundfont	string	Soundfont used for this variant
aug_room_type	string	Simulated room (studio, gym, garage, etc.)
aug_room_wet_dry	float	Reverb wet/dry mix
aug_mic_distance	float	Simulated mic distance (meters)
aug_mic_type	string	Mic model (condenser, dynamic)
aug_compression_ratio	float	Audio compression ratio
aug_noise_level_db	float	Background noise level
aug_bit_depth	int	Bit depth (if degraded)
aug_sample_rate	int	Sample rate (if degraded)

Exercise-Level Scores (0-100)

Column	Type	Description
timing_accuracy	float	How close strokes are to the ideal beat grid
timing_consistency	float	Variability of timing deviations
tempo_stability	float	Drift from target tempo over time
subdivision_evenness	float	Evenness of subdivisions within beats
velocity_control	float	Consistency of stroke dynamics
accent_differentiation	float	Contrast between accented and unaccented strokes
accent_accuracy	float	Whether accents land on the correct beats
hand_balance	float	Evenness between left and right hands
weak_hand_index	float	Relative weakness of the non-dominant hand
flam_quality	float	Quality of flam rudiment execution
diddle_quality	float	Quality of diddle/double-stroke execution
roll_sustain	float	Sustain quality for roll rudiments
groove_feel_proxy	float	Overall rhythmic feel metric
overall_score	float	Weighted composite score

The 40 PAS Rudiments

Single stroke roll, single stroke four, single stroke seven, multiple bounce 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, flam, flam accent, flam tap, flamacue, flam paradiddle, single flammed mill, flam paradiddle-diddle, pataflafla, swiss army triplet, inverted flam tap, flam drag, drag, single drag tap, double drag tap, lesson 25, single dragadiddle, drag paradiddle #1, drag paradiddle #2, single ratamacue, double ratamacue, triple ratamacue, single paradiddle, double paradiddle, triple paradiddle, paradiddle-diddle.

Generation

This dataset was generated using SOUSA, which synthesizes realistic drum performances by:

1. Defining rudiment patterns from YAML specifications
2. Creating player profiles with correlated skill dimensions
3. Generating MIDI with profile-based timing/velocity variation
4. Rendering audio via FluidSynth with multiple soundfonts
5. Applying augmentation chains (room IR, mic modeling, compression, noise)
6. Computing hierarchical quality labels at stroke, measure, and exercise level

pip install sousa
python3 scripts/generate_dataset.py --preset full --with-audio

Citation

@dataset{keown2026sousa,
  title={SOUSA: Synthetic Open Unified Snare Assessment},
  author={Keown, Zak},
  year={2026},
  url={https://huggingface.co/datasets/zkeown/sousa},
  publisher={Hugging Face}
}