Slakh2100: 2,100 Synthesized Multi-Track Music Pieces with Stems

Slakh2100 is a large-scale dataset containing 2,100 automatically mixed music tracks with isolated instrument stems and aligned MIDI data. Created by Manilow et al. in 2019 at Northwestern University, it is designed for music information retrieval and source separation research. The dataset is hosted by schism-audio on Hugging Face.

Use Cases

• Train music source separation models based on the isolated instrument stems.
• Develop automatic music transcription systems using the aligned MIDI data.
• Benchmark multi-track music generation models based on the synthesized tracks.
• Research music structure analysis based on the multi-track format.

Strengths

• Contains 2,100 music tracks, providing a substantial corpus for training.
• Includes both isolated audio stems and aligned MIDI data, offering multimodal ground truth.
• Created by a research team at Northwestern University, suggesting an academic foundation.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Northwestern University (Manilow et al., 2019)

Collection Method

Automatically synthesized and mixed.

Freshness

Last updated 2026-03-10 22:05:44; freshness should be verified.