Slakh2100

Quick Start

from huggingface_hub import snapshot_download

# Download the full dataset (~500GB extracted)
path = snapshot_download(
    repo_id="schismaudio/slakh2100",
    repo_type="dataset",
)

Dataset Description

Slakh2100 (Synthesized Lakh) is a large-scale multi-track music dataset containing 2,100 automatically mixed tracks with isolated instrument stems and aligned MIDI. Created by Manilow et al. (2019) at Northwestern University, Slakh uses MIDI files from the Lakh MIDI Dataset rendered through professional-grade VST instruments to produce realistic multi-track audio.

Each track contains isolated stems for every instrument (drums, bass, guitar, piano, strings, etc.), making Slakh the standard benchmark dataset for music source separation research. The "redux" version hosted here uses lossless FLAC compression to reduce the download size from the original ~500GB WAV to ~105GB.

Dataset Structure

Directory Layout

slakh2100_flac_redux/
  Track00001/
    mix.flac                  # Full mix of all instruments
    stems/
      S00.flac               # Stem 0 (e.g., drums)
      S01.flac               # Stem 1 (e.g., bass)
      S02.flac               # Stem 2 (e.g., piano)
      ...
    MIDI/
      S00.mid                # MIDI for stem 0
      S01.mid                # MIDI for stem 1
      ...
    metadata.yaml            # Instrument labels, plugin info, mix parameters
  Track00002/
    ...

Data Fields

Field	Type	Description
mix.flac	Audio	Full mix of all instrument stems (44.1kHz stereo FLAC)
stems/S*.flac	Audio	Isolated instrument stems (44.1kHz stereo FLAC)
MIDI/S*.mid	MIDI	Aligned MIDI file for each stem
metadata.yaml	YAML	Instrument labels, VST plugin names, mixing parameters

Data Splits

Split	Tracks	Description
train	1,500	Training set
validation	375	Validation set
test	225	Test set

Splits are defined by directory structure. Each split contains non-overlapping tracks.

Audio Format

• Sample rate: 44.1 kHz
• Channels: Stereo
• Format: FLAC (lossless compression)
• Average stems per track: ~5-10 (varies by arrangement complexity)

Instrument Taxonomy

Slakh maps 128 General MIDI instrument programs to 34 instrument classes. Common classes include:

Class	Instruments
Drums	All MIDI percussion (channel 10)
Bass	Acoustic Bass, Electric Bass, Synth Bass
Piano	Acoustic Piano, Electric Piano
Guitar	Acoustic Guitar, Electric Guitar (clean/distorted)
Strings	Violin, Viola, Cello, String Ensemble
Brass	Trumpet, Trombone, French Horn
Reed	Saxophone, Clarinet, Oboe
Synth Lead	Various synthesizer lead patches
Synth Pad	Various synthesizer pad patches

The full mapping is defined in each track's metadata.yaml.

Usage Examples

Load a track and its stems

import soundfile as sf
import yaml
from pathlib import Path

track_dir = Path("slakh2100_flac_redux/Track00001")

# Load the full mix
mix, sr = sf.read(track_dir / "mix.flac")

# Load metadata to identify instruments
with open(track_dir / "metadata.yaml") as f:
    meta = yaml.safe_load(f)

# Load individual stems
for stem_name, stem_info in meta["stems"].items():
    audio, sr = sf.read(track_dir / "stems" / f"{stem_name}.flac")
    print(f"{stem_name}: {stem_info['inst_class']} ({audio.shape[0] / sr:.1f}s)")

Extract drum stems for transcription

import yaml
import soundfile as sf
from pathlib import Path

slakh_root = Path("slakh2100_flac_redux")

for track_dir in sorted(slakh_root.glob("Track*")):
    with open(track_dir / "metadata.yaml") as f:
        meta = yaml.safe_load(f)
    for stem_name, stem_info in meta["stems"].items():
        if stem_info["inst_class"] == "Drums":
            audio, sr = sf.read(track_dir / "stems" / f"{stem_name}.flac")
            midi_path = track_dir / "MIDI" / f"{stem_name}.mid"
            print(f"{track_dir.name}/{stem_name}: drums ({audio.shape[0] / sr:.1f}s)")

Dataset Creation

Source Data

Slakh is built from the Lakh MIDI Dataset (LMD), a collection of 176,581 unique MIDI files. The 2,100 tracks in Slakh were selected from LMD to maximize instrument diversity and arrangement complexity. Each MIDI file was rendered through professional-grade VST instruments using a custom synthesis pipeline.

Synthesis Pipeline

MIDI tracks were assigned to specific VST instruments based on their General MIDI program numbers. The rendering pipeline uses high-quality commercial and open-source VST plugins to produce audio that sounds significantly more realistic than General MIDI synthesis. Each instrument is rendered as an isolated stem, then mixed together to create the full track.

Annotations

MIDI files serve as ground-truth annotations. Since the audio is synthesized directly from MIDI, the alignment between audio and MIDI is exact (sample-accurate). No manual annotation was needed.

Known Limitations

• Synthesized audio only: Despite using professional VST instruments, the audio is synthesized from MIDI. Real recordings have room acoustics, bleed between instruments, and performance nuances not captured by MIDI.
• MIDI-derived performances: Dynamics and timing are limited to what MIDI captures. Subtle expressive details (ghost notes, brush work, muting techniques) may be missing or simplified.
• Pop/rock bias: The Lakh MIDI Dataset skews toward popular Western music genres. Underrepresented genres include jazz, classical, and non-Western music.
• Variable stem quality: Some VST renderings are more realistic than others. Certain instrument classes (e.g., acoustic guitar, vocals) are particularly difficult to synthesize convincingly.
• No vocals: The dataset does not contain vocal stems, as MIDI cannot represent singing.

Related Datasets

This dataset is part of the Drum Audio Datasets collection by schismaudio. Related datasets:

• schismaudio/groove-midi-dataset -- 13.6 hours: Real drum performances with aligned MIDI from a Roland TD-11
• schismaudio/e-gmd -- 444 hours: GMD performances re-rendered with 43 different drum kits
• schismaudio/stemgmd -- 1,224 hours: Isolated per-instrument drum stems from 10 acoustic kits
• schismaudio/star-drums -- Real-world drum recordings with MIDI annotations

Citation

@inproceedings{manilow2019cutting,
  title={Cutting Music Source Separation Some Slakh: A Dataset to Study the Impact of Training Data on Accuracy},
  author={Manilow, Ethan and Wichern, Gordon and Seetharaman, Prem and Le Roux, Jonathan},
  booktitle={Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
  year={2019}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-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 (Manilow et al., Northwestern University).