Musdb Eval

musdb-eval

Hybrid Transformers for Music Source Separation — Rouard et al. (2022) (arXiv:2211.08553, 2022)

What this evaluates

Evaluates a model's ability to isolate individual musical stems (vocals, drums, bass, other) from mixed audio recordings, testing long-range context modeling and cross-domain attention capabilities in source separation.

Datasets

• MUSDB — total ?; splits: test (-1)

Metrics

• SDR (primary) — range: dB
  • Signal-to-Distortion Ratio measures the ratio of the target source energy to the sum of interference, artifact, and noise energy. Higher values indicate better separation quality.

Input / output format

Input: Mixed audio waveform containing multiple overlapping musical sources.

Output: Four separated audio stems corresponding to vocals, drums, bass, and other.

Scoring recipe

sdr_scores = []
for source in ['vocals', 'drums', 'bass', 'other']:
    pred = model.predict(mixed_audio)[:, source]
    true = ground_truth[:, source]
    sdr = compute_sdr(pred, true)
    sdr_scores.append(sdr)
return mean(sdr_scores)

Common pitfalls

• SDR values are highly sensitive to phase alignment and amplitude normalization between predicted and ground truth stems; inconsistent preprocessing leads to non-comparable scores.
• Different source separation papers sometimes report SDR on different subsets of MUSDB or with different reference signal alignments, making direct benchmark comparisons difficult.

Evidence (verbatim from paper)

reach state-of-the-art 9.20 dB SDR on MUSDB.

Citation

@misc{rouard2022hybrid,
  title={Hybrid Transformers for Music Source Separation},
  author={Rouard et al. (2022)},
  year={2022},
  note={arXiv:2211.08553}
}

• arXiv: 2211.08553