Slakh2100: A Dataset for Music Source Separation Research

Slakh2100 is a dataset created to study the impact of training data quality and quantity on music source separation. The dataset is a mirror copy hosted on Hugging Face, with its official repository on Zenodo. It was introduced in a 2019 IEEE WASPAA paper by researchers including Ethan Manilow.

Use Cases

• Training source separation models based on the dataset's multi-track music data.
• Benchmarking algorithm performance based on the dataset's structured audio sources.
• Studying the impact of training data quality on separation tasks as referenced in the description.
• Developing tools for automatic music transcription based on isolated instrument stems.

Strengths

• Created for a specific research purpose, as described in a 2019 IEEE conference paper.
• Hosted on multiple platforms (Hugging Face, Zenodo) for accessibility.

Limitations

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

Provenance

Source

Official Slakh2100 repo on Zenodo, as referenced.

Collection Method

Likely generated or synthesized for music source separation research, as inferred from the paper title.

Time Range

Associated with a 2019 research publication.

Freshness

Last updated 2026-04-25 05:43:44; freshness should be verified.

Geography

null