Datasets:
license: cc-by-4.0
pretty_name: StemGMD
task_categories:
- audio-to-audio
tags:
- audio
- music
- drums
- source-separation
- drum-source-separation
- drum-demixing
- isolated-stems
- groove-midi-dataset
- music-information-retrieval
configs:
- config_name: archives
data_files:
- split: train
path: metadata.csv
StemGMD
This repository mirrors the StemGMD archive files released on Zenodo as "StemGMD: A Large-Scale Audio Dataset of Isolated Drum Stems for Deep Drums Demixing".
StemGMD builds on Magenta's Groove MIDI Dataset. The upstream authors mapped the original 22 MIDI pitches to nine canonical drum instruments and synthesized isolated 16-bit/44.1 kHz stereo WAV stems with ten Logic Pro X acoustic drum kits. The release contains 1224 hours of audio, corresponding to more than 136 hours of full-kit mixtures, plus single hits at ten velocities. The fully extracted dataset requires just over 1.13 TB.
This mirror preserves the upstream archive files and provides a metadata-only dataset viewer manifest. It intentionally does not expand every WAV file into the viewer because the dataset is very large.
Source Records
- Part 1: https://zenodo.org/records/7860223
- Part 2: https://zenodo.org/records/7882857
- License: Creative Commons Attribution 4.0 International
- DOI:
10.5281/zenodo.7860223and10.5281/zenodo.7882857
Files
See metadata.csv for archive names, sizes, MD5 checksums, source URLs, and paths in this repository.
Reconstructing the Dataset
Download all files under archives/ into one directory, then run the upstream unzip_StemGMD.sh script from that directory. The extracted dataset is expected to require just over 1.13 TB.
Citation
@article{mezza2024,
title = {Toward deep drum source separation},
author = {Alessandro Ilic Mezza and Riccardo Giampiccolo and Alberto Bernardini and Augusto Sarti},
journal = {Pattern Recognition Letters},
volume = {183},
pages = {86-91},
year = {2024},
issn = {0167-8655},
doi = {https://doi.org/10.1016/j.patrec.2024.04.026}
}