schism-audio/groove-midi-dataset

Groove MIDI Dataset

Quick Start

from datasets import load_dataset

# Loads the 1,090 audio rows from metadata.csv using the AudioFolder convention.
audio_rows = load_dataset("schism-audio/groove-midi-dataset", split="train")

# The official GMD split is preserved as a column.
official_train = audio_rows.filter(lambda row: row["split"] == "train")

# Download a corresponding MIDI file.
from huggingface_hub import hf_hub_download

first = audio_rows[0]
midi_path = hf_hub_download(
    repo_id="schism-audio/groove-midi-dataset",
    repo_type="dataset",
    filename=first["midi_path"],
)

Dataset Description

The Groove MIDI Dataset (GMD) is a collection of 13.6 hours of aligned MIDI and synthesized audio recordings of drum performances by 10 drummers (80%+ professional). Created by Google Magenta, it contains 1,150 MIDI files covering over 22,000 measures across 18 musical styles.

Each performance was recorded on a Roland TD-11 electronic drum kit, capturing both expressive MIDI (with velocity and precise timing) and synthesized audio output. The dataset is widely used for research in drum transcription, groove modeling, humanization, and rhythmic generation.

GMD is the foundation for several derivative datasets, including E-GMD (re-rendered with 43 kits) and StemGMD (multi-kit isolated stems).

This repository mirrors the official Magenta MIDI metadata exactly: info.csv has 1,150 rows and the uploaded MIDI files match the official groove-v1.0.0-midionly.zip contents. The root metadata.csv follows the Hugging Face AudioFolder convention with a file_name column for the 1,090 WAV files currently present in this repository. The remaining 60 MIDI performances listed in info.csv do not have corresponding uploaded audio files.

Dataset Structure

Data Fields

The default Hugging Face dataset view is built from metadata.csv and includes the audio rows that have uploaded WAV files:

Field	Type	Description
audio	Audio	Decoded WAV audio referenced by metadata.csv/file_name
drummer	string	Anonymized drummer ID (drummer1–drummer10)
session	string	Recording session (e.g., session1, eval_session)
id	string	Unique performance identifier
style	string	Musical genre (18 styles)
bpm	int	Tempo in beats per minute
beat_type	string	"beat" (steady pattern) or "fill" (fill passage)
time_signature	string	Time signature (e.g., 4-4, 6-8)
midi_path	string	Relative path to the corresponding MIDI file
duration	float	Duration in seconds
split	string	Official GMD split: train, validation, or test

The canonical full MIDI metadata remains available in info.csv:

Field	Type	Description
midi_filename	string	Relative path to the MIDI file
audio_filename	string	Relative path to the WAV audio file (22050 Hz, mono) when present
drummer	string	Anonymized drummer ID (drummer1–drummer10)
session	string	Recording session (e.g., session1, eval_session)
id	string	Unique performance identifier
style	string	Musical genre (18 styles)
bpm	int	Tempo in beats per minute
beat_type	string	"beat" (steady pattern) or "fill" (fill passage)
time_signature	string	Time signature (e.g., 4-4, 6-8)
duration	float	Duration in seconds
split	string	Official GMD split: train, validation, or test

Data Splits

Split	MIDI / metadata rows	WAV files present	Beats	Fills	Duration
train	897	846	378	519	648.5 min
validation	124	120	48	76	82.2 min
test	129	124	77	52	84.3 min
Total	1,150	1,090	503	647	814.9 min (13.6 h)

The Hugging Face Dataset Viewer uses metadata.csv, so it previews the 1,090 audio rows with actual uploaded WAV files. The official GMD split is stored in the split column; the Viewer may expose these rows under a generated train split because AudioFolder treats the root metadata.csv as one loadable audio table. Use info.csv when you need the complete 1,150-row MIDI inventory.

Class Taxonomy

The dataset uses 9 canonical drum instrument classes, mapped from 22 raw MIDI pitches on the Roland TD-11:

Class	Instrument	MIDI Pitches
0	Kick	36
1	Snare (Head)	38
2	Snare (Rim)	37, 40
3	Closed Hi-Hat	42, 22
4	Open Hi-Hat	46, 26
5	Low Tom	43, 58
6	Mid Tom	47, 45
7	High Tom	50, 48
8	Ride / Crash	49, 51, 55, 57, 52

Usage Examples

Parse MIDI onsets

import pretty_midi

midi = pretty_midi.PrettyMIDI("path/to/file.mid")
for instrument in midi.instruments:
    for note in instrument.notes:
        print(f"Time: {note.start:.3f}s, Pitch: {note.pitch}, Velocity: {note.velocity}")

Load metadata

import pandas as pd

info = pd.read_csv("info.csv")
# Filter by style
rock = info[info["style"] == "rock"]
print(f"Rock tracks: {len(rock)}, Total duration: {rock['duration'].sum():.0f}s")

Dataset Creation

Source Data

All performances were recorded at Magenta's recording studio by 10 professional and semi-professional drummers playing on a Roland TD-11 electronic drum kit. Drummers performed along with a metronome at specified tempos across 18 musical styles.

Annotations

Annotations are the raw MIDI output from the electronic drum kit — no manual annotation was needed. MIDI captures exact onset times, pitches, and velocities directly from the pad triggers. Audio was synthesized by the TD-11's internal sound engine.

Known Limitations

• Synthesized audio only: Audio is the TD-11's internal sound engine output, not acoustic drums. This limits generalization to real-world acoustic drum sounds.
• Single kit: All recordings use one drum kit (Roland TD-11), so the model may overfit to this kit's timbral characteristics.
• Limited timbral diversity: To address this, consider using E-GMD (43 kits) or StemGMD (10 acoustic kits).
• Electronic pad dynamics: Velocity response differs from acoustic drums — pad triggers have fixed dynamic curves.

Related Datasets

This dataset is part of the Drum Audio Datasets collection by schism-audio. Related datasets:

• schism-audio/e-gmd — 444 hours: E-GMD re-renders GMD performances with 43 different kits
• schism-audio/stemgmd — 1,224 hours: StemGMD provides isolated per-instrument stems from 10 acoustic kits
• schism-audio/star-drums — Real-world drum recordings with accompaniment

Citation

@inproceedings{gillick2019learning,
  title={Learning to Groove with Inverse Sequence Transformations},
  author={Gillick, Jon and Roberts, Adam and Engel, Jesse and Eck, Douglas and Bamman, David},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2019}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0).

You are free to share and adapt this dataset for any purpose, including commercial use, as long as you give appropriate credit to the original authors (Google Magenta).