Expanded Groove MIDI Dataset — Augmented (E-GMD-Aug)

Quick Start

import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download

# Download a single shard
path = hf_hub_download(
    "schismaudio/e-gmd-aug",
    filename="features/train-00000.parquet",
    repo_type="dataset",
)
table = pq.read_table(path)
print(f"Rows: {table.num_rows}, Columns: {table.column_names}")

Dataset Description

E-GMD-Aug is an augmented derivative of the Expanded Groove MIDI Dataset (E-GMD), designed for training automatic drum transcription (ADT) models that generalize from synthesized to real-world acoustic drum recordings.

The original E-GMD contains ~49,000 VST-rendered drum recordings that sound clean and synthetic. Models trained on this data suffer a domain gap when applied to real-world recordings with room acoustics, microphone coloration, and background noise. E-GMD-Aug addresses this by applying three waveform-level augmentations to each training track before computing mel spectrograms:

1. Room Impulse Response (RIR) convolution — Convolves dry audio with real and simulated RIRs from OpenSLR-28 (60,000+ RIRs), simulating diverse room acoustics.
2. Parametric EQ — Applies random low shelf, mid peak, and high shelf filters, simulating microphone and mixing coloration.
3. Background noise mixing — Mixes in point-source noise recordings at 20–40 dB SNR, simulating ambient room noise.

Each training track produces 1 dry (unaugmented) copy + 3 augmented copies (4x data multiplier). Augmentation parameters are randomized per copy and stored in the augmentation column for reproducibility.

This dataset contains pre-computed features (mel spectrograms + onset/velocity targets), not raw audio. It is designed for direct use with DrumscribbleCNN.

Dataset Structure

Data Fields

Field	Type	Description
mel_spectrogram	binary	128-band mel spectrogram, float32 (128 × n_frames)
onset_targets	binary	Onset target matrix, float32 (26 × n_frames)
velocity_targets	binary	Velocity target matrix, float32 (26 × n_frames)
n_frames	int64	Number of time frames
n_mels	int64	Number of mel bands (128)
n_classes	int64	Number of instrument classes (26)
sample_rate	int64	Audio sample rate used (16000)
hop_length	int64	STFT hop length (256)
fps	float64	Frames per second (62.5)
duration	float64	Duration in seconds
split	string	Always "train" (val/test are unaugmented)
augmentation	string	Empty for dry copy; JSON with augmentation params for augmented copies
source_audio	string	Original audio filename
style	string	Musical style (e.g. rock, funk, jazz)
bpm	float64	Tempo in BPM
drummer	string	Drummer ID
session	string	Recording session
beat_type	string	"beat" or "fill"
time_signature	string	Time signature (e.g. 4-4)
kit_name	string	VST drum kit name
source_id	string	Source performance ID

Data Splits

Split	Entries	Rows (1 dry + 3 aug)	Shards
train	35,217	140,868	410

Only the train split is augmented. Validation and test splits are served unaugmented from the original schismaudio/e-gmd repo.

File Layout

features/
  train-00000.parquet
  train-00001.parquet
  ...
  train-00409.parquet

Augmentation Parameters

Each augmented row stores its parameters in the augmentation column as JSON:

{
  "rir_idx": 42531,
  "wet_mix": 0.65,
  "low_shelf_db": -3.21,
  "high_shelf_db": 2.45,
  "low_shelf_freq": 125.0,
  "high_shelf_freq": 6200.0,
  "mid_freq": 1250.0,
  "mid_db": 1.85,
  "mid_q": 1.2,
  "noise_idx": 312,
  "snr_db": 28.5
}

Dry (unaugmented) copies have an empty string in the augmentation column.

Augmentation Details

RIR Convolution

RIRs are sourced from schismaudio/openslr-rirs (OpenSLR-28):

• ~60,000 simulated RIRs across diverse room geometries
• 417 real isotropic RIRs recorded in actual rooms
• Wet/dry mix ratio randomized between 0.3–0.9

Parametric EQ

Three-band parametric equalizer:

• Low shelf: 80–200 Hz, ±6 dB
• Mid peak: 300–3000 Hz, ±4 dB, Q 0.7–2.0
• High shelf: 4000–8000 Hz, ±6 dB

Background Noise

Point-source noise recordings from OpenSLR-28 (843 noise WAVs):

• SNR randomized between 20–40 dB
• Noise is looped if shorter than the audio

Usage with DrumscribbleCNN

from drumscribble.data.features import ParquetFeaturesDataset
from huggingface_hub import HfApi, hf_hub_download

# Download all train shards
api = HfApi()
files = [f for f in api.list_repo_files("schismaudio/e-gmd-aug", repo_type="dataset")
         if f.startswith("features/train-") and f.endswith(".parquet")]
paths = [hf_hub_download("schismaudio/e-gmd-aug", f, repo_type="dataset") for f in files]

# Create dataset (lazy loading — one shard in memory at a time)
dataset = ParquetFeaturesDataset(paths, chunk_frames=625)
print(f"Training chunks: {len(dataset):,}")

Dataset Creation

Generated by compute_features_aug.py, a PEP 723 UV script run as an HF Job on L4 GPU hardware. The pipeline:

1. Downloads E-GMD raw audio from Google Cloud Storage (~90 GB)
2. Downloads OpenSLR RIRs from HF Hub
3. For each training track: computes 1 dry + 3 augmented mel spectrograms
4. Streams sharded Parquet files to HF Hub (delete-after-upload to manage disk)

Related Datasets

• schismaudio/e-gmd — Original E-GMD with raw audio + pre-computed features (unaugmented)
• schismaudio/star-drums-aug — Augmented STAR dataset (same pipeline)
• schismaudio/openslr-rirs — Room impulse responses used for augmentation

Citation

@article{callender2020improving,
  title={Improving Perceptual Quality of Drum Transcription with the Expanded Groove MIDI Dataset},
  author={Callender, Lee and Hawthorne, Curtis and Engel, Jesse},
  journal={arXiv preprint arXiv:2004.00188},
  year={2020}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0), the same license as the original E-GMD.