MUSDB18-HQ: 150 Uncompressed Music Tracks for Source Separation

MUSDB18-HQ is an uncompressed audio dataset containing 150 full-track songs across different styles, created by Zafar Rafii et al. in 2019. It provides stereo mixtures and isolated sources (vocals, bass, drums, other) for 100 training and 50 test songs, encoded as 44.1kHz WAV files. The dataset serves as a reference for designing and evaluating source separation algorithms and was used in the SiSEC 2018 campaign.

Use Cases

• Training supervised source separation models based on the provided 100-song training set with isolated stems.
• Benchmarking separation algorithm performance on the 50-song test set against SiSEC 2018 participants.
• Developing karaoke or remix applications by isolating vocal tracks from the provided stereo mixtures.
• Researching music information retrieval techniques using high-fidelity, uncompressed audio sources.

Strengths

• Contains 150 professionally-produced, full-length music tracks.
• Provides ground-truth isolated stems (vocals, bass, drums, other) for all 150 tracks.
• Files are uncompressed, high-fidelity WAV format at 44.1kHz stereo.
• Structured into 100 tracks for training and 50 for testing, supporting standard evaluation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last update date is unknown; freshness unverified.
• License is restrictive for commercial use, requiring permission from multiple copyright holders.

Provenance

Source

Compiled from DSD100, MedleyDB, Native Instruments, and The Easton Ellises sources.

Collection Method

Curated collection of existing multitrack music libraries and contributions.

Time Range

Temporal coverage of the original source recordings is not specified.

Freshness

Dataset created in 2019; last update date is unknown.

Geography

Spatial coverage is not specified; likely international given the source contributors.