StemGMD: A Large-Scale Audio Dataset of Isolated Drum Stems for Deep Drums Demixing - part 1

We introduce StemGMD, a new large-scale dataset of isolated drum stems that builds upon the extensive MIDI collection found in Magenta's Groove MIDI Dataset (GMD).

GMD is a 13.6-hour corpus of expressive drum performances executed by ten drummers on a Roland TD-11 electronic drum kit. It contains 1150 MIDI files along with the corresponding full-kit audio mixtures.

As a first step in creating StemGMD, we mapped the 22 different MIDI pitches found in the original files onto nine canonical instruments through the reduction scheme proposed  in J. Gillick, A. Roberts, J. Engel, D. Eck, and D. Bamman, "Learning to groove with inverse sequence transformations," in International Conference on Machine Learning (ICML), vol. 97, 2019, pp. 2269–2279.

Each of the nine resulting MIDI channels was manually synthesized as a 16-bit/44.1 kHz stereo WAV file using ten realistic-sounding acoustic drum kits sourced from the Logic Pro X sample libraries, i.e., Bluebird, Brooklyn, Detroit Garage, East Bay, Heavy, Motown Revisited, Portland, Retro Rock, Roots, and SoCal.

As a result, StemGMD contains 1224 hours of audio, which correspond to more than 136 hours of full-kit mixtures. Moreover, StemGMD also contains single hits for each of the drum pieces at ten different velocities ranging from 30 to 127.

To the best of our knowledge, StemGMD is the largest publicly available dataset of drums to date. Moreover, it is the first collection of single-instrument clips from all nine pieces in a canonical drum kit, making it well-suited for training deep drums demixing models.

*** THIS IS PART 1 OF 2 ***

Download part 2 here: https://zenodo.org/records/7882857  (now available!) 

After downloading both parts, run unzip_StemGMD.sh to build the dataset from the split archive files.
Once unzipped, StemGMD will take just over 1.13 TB of memory.

The dataset is made available under a Creative Commons Attribution 4.0 International (CC BY 4.0) License.

____________________________

We employed the dataset in our paper titled "Toward Deep Drum Source Separation," published in Pattern Recognition Letters. 

@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}
}