Expanded Groove MIDI Dataset (E-GMD)

This repository mirrors version 1.0.0 of Google's Expanded Groove MIDI Dataset (E-GMD) for access through the Hugging Face Hub.

E-GMD is a large dataset of human drum performances with audio recordings annotated in MIDI. It contains 444.5 hours of audio from 43 drum kits, with the same train, validation, and test split definitions as the original Groove MIDI Dataset.

Quick Start

from datasets import load_dataset

ds = load_dataset("schism-audio/e-gmd", split="train", streaming=True)
first = next(iter(ds))
print(first["audio"], first["midi_path"], first["split"])

Repository Layout

The original archive contains some session folders with more than 10,000 files, which exceeds Hugging Face Hub's per-folder repository limit. This mirror keeps the original filenames and original drummer/session paths, but stages files under split and kit directories:

audio/{split}/{kit_slug}/{original_drummer/session/path}.wav
midi/{split}/{kit_slug}/{original_drummer/session/path}.midi
metadata.csv
metadata/{split}.csv
metadata/all.csv
e-gmd-v1.0.0.csv

The root metadata.csv follows the AudioFolder convention and contains one row per WAV file. The file_name column points at the audio file, and the remaining columns preserve E-GMD metadata plus the paired MIDI path.

The metadata/*.csv files are retained from the original mirror and add these path columns:

• file_name: audio path relative to the split folder
• audio_path: audio path relative to the repository root
• midi_path: paired MIDI path relative to the repository root
• original_audio_filename: original archive audio path
• original_midi_filename: original archive MIDI path

Loading Audio With Metadata

The default dataset is loadable with datasets through the AudioFolder convention:

from datasets import load_dataset

train = load_dataset("schism-audio/e-gmd", split="train", streaming=True)
print(train[0]["audio"], train[0]["midi_path"])

The official train/test/validation split is preserved in the split column. The paired MIDI file for each row is available through midi_path.

Splits

License

The dataset is made available by Google LLC under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

Source

• Official Magenta page: https://magenta.tensorflow.org/datasets/e-gmd
• Zenodo record: https://zenodo.org/records/4300943
• Version: 1.0.0
• Original full archive SHA256: 7d9a264fb4c9eabd9fec09d5f8e333192f529b1a1b845d170279a977ac436053

Citation

If you use this dataset, cite the original E-GMD paper and specify version 1.0.0:

@misc{callender2020improving,
    title={Improving Perceptual Quality of Drum Transcription with the Expanded Groove MIDI Dataset},
    author={Lee Callender and Curtis Hawthorne and Jesse Engel},
    year={2020},
    eprint={2004.00188},
    archivePrefix={arXiv},
    primaryClass={cs.SD}
}

📊 Structured Schema (Zero-Fabrication)

Estimated Rows: 45,537