Musdb18 Hq Eval

musdb18-hq-eval

Discrete Token Modeling for Multi-Stem Music Source Separation with Language Models — Lyu et al. (2026) (arXiv:2604.09371, 2026)

What this evaluates

Evaluates the perceptual quality of multi-stem music source separation (vocals, drums, bass, other) generated by a discrete token modeling framework. It measures how well the model separates audio tracks compared to discriminative baselines, focusing on perceptual audio quality and vocal intelligibility/naturalness.

Datasets

• MUSDB18-HQ — total 50; splits: test (50)

Metrics

• ViSQOL (primary) — range: [1, 5]
  • Estimates perceptual audio quality on a scale of 1 to 5 using spectro-temporal representations with internal patch alignment to tolerate minor temporal shifts.
• DNSMOS — range: [0, 5]
  • Reports three sub-scores for vocals: signal quality (SIG), background noise quality (BAK), and overall quality (OVRL).
• NISQA — range: [0, 5]
  • Provides a non-intrusive overall quality estimate as a proxy for perceived vocal quality.

Input / output format

Input: 4.0-second mono audio segments at 48 kHz, containing mixed multi-track music (vocals, drums, bass, other).

Output: Separated audio tracks for vocals, drums, bass, and other, decoded from discrete tokens via a neural audio codec (HCodec).

Scoring recipe

# Encode and decode both reference and predicted tracks through HCodec to remove codec reconstruction influence
ref_decoded = HCodec.decode(HCodec.encode(reference))
pred_decoded = HCodec.decode(HCodec.encode(prediction))

# Compute ViSQOL for each track and average
visqol_scores = [ViSQOL(ref_decoded[t], pred_decoded[t]) for t in ['vocals', 'drums', 'bass', 'other']]
avg_visqol = mean(visqol_scores)

# Compute DNSMOS and NISQA for vocals track
dnsmos = DNSMOS(ref_decoded['vocals'], pred_decoded['vocals'])
nisqa = NISQA(ref_decoded['vocals'], pred_decoded['vocals'])

Common pitfalls

• Using sample-aligned metrics like SI-SNR or PESQ, which are unsuitable due to autoregressive decoding causing minor temporal shifts.
• Evaluating raw codec outputs without re-encoding/decoding through HCodec, which would conflate separation quality with codec reconstruction artifacts.
• Ignoring track-specific evaluation and only reporting overall averages without separating vocals, drums, bass, and other.

Evidence (verbatim from paper)

For overall separation quality, we use ViSQOL (Virtual Speech Quality Objective Listener) [hines2015visqol, chinen2020visqol], which estimates perceptual audio quality on a scale of 1 to 5. ViSQOL performs internal patch alignment in the spectro-temporal domain, making it tolerant to the small timing variations introduced by autoregressive decoding. ... To remove the influence of codec reconstruction, both the reference signals and the outputs of all baseline models are encoded and decoded through HCodec before evaluation.

Citation

@misc{lyu2026discrete,
  title={Discrete Token Modeling for Multi-Stem Music Source Separation with Language Models},
  author={Lyu et al. (2026)},
  year={2026},
  note={arXiv:2604.09371}
}

• arXiv: 2604.09371