← Glossary

MUSDB18

A benchmark dataset of 150 songs with separate vocals, drums, bass, and other stems, widely used to train and evaluate music separation models. Restricted to educational and noncommercial use.

MUSDB18 contains 150 full-length songs, totaling approximately ten hours of music, with separate vocals, drums, bass, and “other” stems. Created in 2017 by researchers including AudioShake's Head of Research, Fabian-Robert Stöter, it became a standard benchmark for music source separation. The recordings were provided for educational and research purposes; commercial use requires appropriate rights-holder permission. MUSDB18-HQ is the uncompressed version. Use of this dataset in model training can raise separate questions about the commercial use of resulting model weights.

Links: MUSDB18 · MUSDB18-HQ · Understanding MUSDB18, model weights, and licensing

Where this comes up at AudioShake
No items found.