---
license: cc-by-4.0
task_categories:
  - audio-classification
tags:
  - drums
  - percussion
  - drum-transcription
  - source-separation
  - midi
  - stems
  - multi-kit
  - acoustic-drums
pretty_name: "StemGMD"
size_categories:
  - 10K<n<100K
---

# StemGMD

## Quick Start

```python
from huggingface_hub import snapshot_download

# Download the full dataset (~1.13TB extracted)
path = snapshot_download(
    repo_id="schismaudio/stemgmd",
    repo_type="dataset",
)
```

## Dataset Description

**StemGMD** is a large-scale multi-kit drum dataset providing 1,224 hours of isolated per-instrument drum stems with aligned MIDI annotations. Created by Ferroni et al. (2023), StemGMD takes the same 1,150 MIDI performances from the [Groove MIDI Dataset (GMD)](https://huggingface.co/datasets/schismaudio/groove-midi-dataset) and renders each one through **10 professional acoustic drum kits** using high-quality sample libraries.

Unlike [E-GMD](https://huggingface.co/datasets/schismaudio/e-gmd) (which provides a single mixed drum audio per kit), StemGMD provides **isolated per-instrument stems** -- kick, snare, hi-hat, toms, and cymbals are each rendered as separate audio files. This makes StemGMD uniquely valuable for both drum transcription and drum source separation research, enabling models to learn from clean, instrument-level supervision.

## Dataset Structure

### Directory Layout

```
StemGMD/
  drummer1/
    session1/
      1_rock_80_beat_4-4/
        kit_00/
          kick.wav
          snare.wav
          hihat.wav
          tom_low.wav
          tom_mid.wav
          tom_high.wav
          crash.wav
          ride.wav
          mix.wav
        kit_01/
          ...
        ...
        kit_09/
          ...
        midi/
          file.mid
  drummer2/
    ...
  eval_session/
    ...
```

### Data Fields

| Field | Type | Description |
|-------|------|-------------|
| `kick.wav` | Audio | Isolated kick drum stem (44.1kHz WAV) |
| `snare.wav` | Audio | Isolated snare drum stem |
| `hihat.wav` | Audio | Isolated hi-hat stem (open + closed) |
| `tom_low.wav` | Audio | Isolated low tom stem |
| `tom_mid.wav` | Audio | Isolated mid tom stem |
| `tom_high.wav` | Audio | Isolated high tom stem |
| `crash.wav` | Audio | Isolated crash cymbal stem |
| `ride.wav` | Audio | Isolated ride cymbal stem |
| `mix.wav` | Audio | Full drum mix (sum of all stems) |
| `file.mid` | MIDI | Ground-truth MIDI with onset times, pitches, and velocities |

### Data Splits

Splits follow the same drummer-based partitioning as GMD (no drummer overlap between splits). Each split is multiplied by 10 kits.

| Split | Performances | Total Recordings | Description |
|-------|-------------|-----------------|-------------|
| `train` | ~800 | ~8,000 | Training set (~800 MIDI x 10 kits) |
| `validation` | ~200 | ~2,000 | Validation set |
| `test` | ~150 | ~1,500 | Test set |

### Audio Format

- **Sample rate**: 44.1 kHz
- **Format**: WAV (uncompressed)
- **Stems per kit rendering**: Up to 9 (8 instrument stems + 1 mix)

## Drum Kits

StemGMD uses 10 professional acoustic drum sample libraries spanning a range of timbral characters:

| Kit ID | Description |
|--------|------------|
| 0--9 | 10 distinct acoustic drum kits from commercial sample libraries |

Each kit provides a different timbral character (bright/dark, tight/open, studio/live), ensuring diverse training data for generalization across drum sounds.

## Usage Examples

### Load isolated stems for one rendering

```python
import soundfile as sf
from pathlib import Path

track_dir = Path("StemGMD/drummer1/session1/1_rock_80_beat_4-4/kit_00")

stems = {}
for stem_file in track_dir.glob("*.wav"):
    audio, sr = sf.read(stem_file)
    stems[stem_file.stem] = audio
    print(f"{stem_file.stem}: {audio.shape[0] / sr:.1f}s, sr={sr}")
```

### Compare the same performance across kits

```python
import soundfile as sf
from pathlib import Path

perf_dir = Path("StemGMD/drummer1/session1/1_rock_80_beat_4-4")

for kit_dir in sorted(perf_dir.glob("kit_*")):
    mix, sr = sf.read(kit_dir / "mix.wav")
    print(f"{kit_dir.name}: {mix.shape[0] / sr:.1f}s, rms={((mix**2).mean()**0.5):.4f}")
```

### Parse MIDI onsets with stem-level alignment

```python
import pretty_midi

midi = pretty_midi.PrettyMIDI("StemGMD/drummer1/session1/1_rock_80_beat_4-4/midi/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}")
```

## Dataset Creation

### Source Data

StemGMD builds on the [Groove MIDI Dataset](https://huggingface.co/datasets/schismaudio/groove-midi-dataset), which contains 1,150 MIDI performances recorded by 10 professional and semi-professional drummers on a Roland TD-11 electronic drum kit at Google Magenta's recording studio.

### Rendering Pipeline

Each MIDI performance was rendered through **10 professional acoustic drum sample libraries**. Rather than rendering a single mixed audio file (as in E-GMD), StemGMD renders each drum instrument as an isolated audio stem. This is accomplished by routing each MIDI pitch class to its corresponding sample in the drum kit and rendering them independently.

All audio was rendered at **44.1kHz** (CD quality), a significant improvement over E-GMD's 16kHz, preserving the full frequency range of cymbals and transients that are critical for drum transcription.

### Annotations

Annotations are inherited directly from GMD: the raw MIDI output from the Roland TD-11 electronic drum kit. Since the audio is synthesized from MIDI, alignment is exact. The same MIDI annotations apply to all 10 kit renderings.

## Known Limitations

- **Synthesized audio only:** Audio is rendered from sample libraries, not recorded from live drumming. While the sample libraries are high-quality, the audio lacks the natural variability of live acoustic recordings (room ambience, microphone bleed, stick-to-head interaction).
- **Sample-based rendering:** Each hit triggers a single sample. Real drums produce continuous timbral variation based on strike position, velocity curves, and sympathetic resonance that sample libraries only partially capture.
- **Fixed MIDI performances:** The underlying 1,150 MIDI performances are identical across all kits. The dataset provides timbral diversity but not additional performance diversity.
- **Electronic pad dynamics:** Velocity values come from Roland TD-11 pad triggers, which have fixed dynamic curves that differ from acoustic drum stick-on-head dynamics.
- **Large size:** At 1.13TB extracted, the dataset requires significant storage. Consider streaming or partial downloads for initial experimentation.

## Related Datasets

This dataset is part of the [Drum Audio Datasets](https://huggingface.co/collections/schismaudio/drum-audio-datasets) collection by [schismaudio](https://huggingface.co/schismaudio). Related datasets:

- [schismaudio/groove-midi-dataset](https://huggingface.co/datasets/schismaudio/groove-midi-dataset) -- 13.6 hours: The original GMD with 1,150 MIDI performances from a single Roland TD-11 kit
- [schismaudio/e-gmd](https://huggingface.co/datasets/schismaudio/e-gmd) -- 444 hours: GMD performances re-rendered with 43 different kits (mixed audio only, no isolated stems)
- [schismaudio/star-drums](https://huggingface.co/datasets/schismaudio/star-drums) -- Real-world drum recordings with MIDI annotations
- [schismaudio/slakh2100](https://huggingface.co/datasets/schismaudio/slakh2100) -- 2,100 multi-track synthesized music with isolated instrument stems

## Citation

```bibtex
@inproceedings{ferroni2023stemgmd,
  title={{StemGMD}: A Large-Scale Multi-Kit Audio Dataset for Automatic Drum Transcription},
  author={Ferroni, Marcello and Lee, Jongho and Fazekas, Gy{\"o}rgy and Sandler, Mark},
  booktitle={Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
  year={2023}
}
```

## License

This dataset is released under the [Creative Commons Attribution 4.0 International License (CC-BY 4.0)](https://creativecommons.org/licenses/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 (Ferroni et al., Queen Mary University of London).