MUSDB-HQ is the standard benchmark dataset for music source separation research

MUSDB-HQ (High Quality) is a dataset of 150 full-length professionally produced songs, each provided as multi-track recordings with separated drum, bass, vocals, and other stems at 44.1 kHz stereo. It is the de facto benchmark for comparing separation models. MUSDB non-HQ (a compressed predecessor) is no longer supported by Demucs training. The test set contains 50 songs withheld from training; models trained on the test set (like mdx_extra) are benchmarked separately and are marked as biased. MUSDB-HQ is hosted on Zenodo under a research-only license.

Examples

Model comparison: models trained without MUSDB test data peak at 8.2 dB SDR (Band-Split RNN); HTDemucs fine-tuned with 800 extra songs also reaches 9.0 dB SDR on the same MUSDB-HQ test set.

Assessment

What is the difference between mdx and mdx_extra in terms of training data? Why is mdx_extra potentially biased for benchmarking purposes?