Source-linked AI summary
Beyond SDR: How Music Source Separation Reshapes Rhythm-Relevant Signal Properties
Chuxin Ding
TL;DR
Music source separation is increasingly used to measure rhythm, but SDR does not establish whether separated stems preserve the attack, envelope, and dynamics that shape perceived timing. This paper benchmarks four open separators on MUSDB18-HQ and finds that onset locations remain robust while transient and dynamic shape acquire model- and input-dependent distortions. The results support reporting separator identity and input conditions rather than relying on SDR alone for rhythmic analysis.
Problem
Music source separation is increasingly used to measure microtiming and dynamics, yet SDR does not reveal whether separation preserves the signal properties those measurements read.
Method
The study benchmarks four open separators across architecture generations on MUSDB18-HQ, measuring onset fidelity, transient and dynamic deviations, model bias, and input-length effects against true stems.
Results
Onset locations remain robust, whereas transient and dynamic shape show systematic model-specific bias that SDR does not predict and are partly perturbed by input length.
Takeaways & Limitations
Rhythmic studies should report the separator name, version, and input conditions because SDR alone cannot determine whether separated dynamics and attacks remain suitable for analysis.
Takeaways & Limitations
The study measures distortion of perceptually relevant properties rather than perception itself, uses a Western pop/rock-centric MUSDB18-HQ corpus, and focuses its analysis on drums.
Abstract
from arXiv · showhide
Music source separation (MSS) is increasingly used not to remix music but to measure it: separated drum stems feed studies of microtiming, dynamics, and groove. The field evaluates separators almost exclusively by signal-to-distortion ratio (SDR), yet microrhythm research shows that a sound's perceived temporal location (its p-centre) is co-determined by its attack and envelope, precisely the properties SDR was not designed to protect. We quantify what four open separators spanning four architecture generations (Spleeter, HT-Demucs, BS-Roformer, SCNetXL) do to rhythm-critical signal properties, using the 50-track MUSDB18-HQ test set, where true stems make every claim falsifiable. Three findings emerge. (1) Onset timing is safe: onset F-measure tracks SI-SDR (Spearman rho = 0.62) and is invariant to input length. (2) Transient and dynamic shape are not: their distortion correlates only weakly with SI-SDR (|rho| <= 0.29), and the model ranking inverts - the SDR leader distorts drum attacks twice as much as its capability-matched CNN counterpart, while the SDR-worst model preserves dynamics better than a mid-pack one. Each model imposes a systematic, model-specific bias on the dynamic profile. (3) Input length reshapes the rendered attack of a fixed passage (marginally more for the transformer; paired p = 0.044) while leaving onset locations untouched. For rhythmic studies, separator choice and input conditions are methodological variables to be reported, and SDR alone cannot stand in for them.
1 Introduction
Music source separation is increasingly used to measure rhythm, but its validity depends on preserving the signal properties those measurements read. This paper tests that assumption with developer and analyst metrics, four open separators, and controlled input lengths.
- SDR measures global reconstruction fidelity but does not specify which rhythm-relevant signal properties absorb distortion.Attack, duration, and envelope co-determine a sound’s perceived temporal location, or p-centre.
- The study turns prior observations of model-specific dynamic reshaping, transient smearing, and input-length dependence into falsifiable claims.
- The evaluation combines developer metrics with onset timing, an eight-descriptor dynamic profile, and ground-truth energy-routing precision.
- Four open separators spanning 2019–2024 architecture generations are benchmarked to test whether SDR predicts analyst-relevant fidelity.
- A controlled input-length experiment tests whether surrounding audio changes transient rendering while preserving onset locations.
- The paper also releases an open, cached, resumable pipeline for reproducing and extending its results.
2 Background and hypotheses
Rhythm studies read event timing, dynamics, and envelope or timbre from separated drum stems, but p-centre evidence makes attack and envelope fidelity part of perceived timing. The paper therefore tests model-specific dynamics bias, input-length sensitivity, and dissociation between SDR and analyst-relevant fidelity.
- Rhythm and groove studies measure onset or microtiming, dynamics, and envelope or timbre from drum stems.
- Attack sharpness and envelope shape can displace perceived temporal location by tens of milliseconds, matching the scale of measured microtiming deviations.
- A1 tests whether separated-stem dynamic profiles deviate from true stems systematically by separator, confounding dynamics statistics with model choice.
- A2 tests whether surrounding context changes rendered transients, dynamics, and labeling, with greater context sensitivity hypothesized for attention models than frame-local CNNs.
- B3 tests whether SDR rankings differ from analyst-relevant fidelity rankings and whether pooled SDR–analyst-distortion correlations are weak.
3 Materials and method
The study evaluates four deterministic open separators on MUSDB18-HQ using developer metrics and analyst-facing measures computed against true stems. It separately benchmarks model and track variation, then sweeps input context around fixed passages to measure descriptor drift.
- Models: Four deterministic open separators emit four stems and include a capability-matched attention/no-attention pair, while proprietary tools are excluded for lacking scriptable open implementations.
- Dataset: The MUSDB18-HQ test subset contains 50 full tracks with true isolated stems, and separations are cached for resumable and re-scorable experiments.
- Metrics: Developer metrics comprise BSS-Eval SDR, SIR, SAR, and ISR plus scale-invariant SDR, recomputed on the study tracks as the comparison baseline.
- Metrics: Analyst metrics measure onset F-measure, an eight-descriptor RMS-envelope dynamic profile, and labeling precision against the true stem.
- Benchmark: The benchmark covers 4 models × 50 tracks × 4 stems, with analysis focused on drums across n = 200 model-track observations.
- Input-length sweep: The input-length sweep separates a fixed 5 s region inside windows of R alone, 15 s, 30 s, and the full track, then measures descriptor drift on R.
- Reproducibility: The implementation records provenance for every separation and isolates GPU and CPU dependency environments on one NVIDIA RTX 4070 SUPER.
4 Results
Across four open separators, SDR aligns with onset timing but not transient or dynamic shape. Separator choice and input context therefore alter rhythm-relevant measurements in model-specific ways that SDR alone does not reveal.
- Developer metrics: 0.05 dB separates the two 2023/24 drum models by median BSS-Eval SDR, making them interchangeable by the field’s metric.Labeling precision remains near-ceiling at 0.94–1.00, leaving energy fidelity as the open question.
- Dynamic fidelity: 1.78 dB versus 4.09 dB median dynamic-variation distortion shows BS-Roformer preserves drum dynamics better than HT-Demucs.HT-Demucs reshapes dynamics about 2.3× more despite 10.1 dB SDR, while Spleeter has 5.0 dB SDR and 2.41 dB distortion.
- Timing versus shape: 0.62 Spearman correlation links onset F-measure with SI-SDR across 200 drum observations, making SDR a reasonable guide to onset placement.The onset relationship is shown in Figure 1(a).
- Timing versus shape: |ρ| ≤ 0.29 characterizes the weak SI-SDR associations with attack-slope, dynamic-variation, and crest distortion.The correlations are −0.17, −0.16, and −0.29 respectively, so SDR cannot stand in for these shape properties.
- Ranking inversions: 1.33 versus 0.64 shows the SDR-leading BS-Roformer distorts drum attacks more than twice as much as capability-matched SCNet-XL.The transient-fidelity ranking therefore reverses at the state-of-the-art frontier.
- Input length: 0.02–0.03 onset-F drift across context windows indicates timing is context-invariant, while transient and dynamic shape drift by an order of magnitude more.Attack rendering is marginally more context-sensitive for the transformer: BS-Roformer exceeds SCNet-XL on 13/20 tracks, with a 1.62× median ratio and p = 0.044.
- Input length: −0.38 and −0.35 mean dynamic-variation changes move CNN and hybrid outputs toward truth with more context, whereas the transformer moves slightly away by +0.09.These deltas are suggestive rather than significant; the transformer’s onset-F nevertheless improves from 0.83 to 0.88.
- Interference versus artifact: SIR and SAR expose an architecture-linked trade-off: BS-Roformer suppresses interference most, while SCNet-XL is most artifact-free despite SDRs within 0.1 dB.Spleeter is worst on both axes and has SAR about 5 dB below modern models.
5 Discussion
For rhythm and groove research, onset locations are comparatively robust, but dynamics and attack shape are model- and input-dependent. Separator choice, version, and input conditions therefore belong in methodological reporting.
- Onset F tracks SDR, reaches 0.89–0.93 for modern drum models, and is invariant to input length.
- Dynamic profiles carry systematic, model-specific bias that SDR does not predict and input length partly perturbs.
- At the current frontier, models differ by 0.05 dB in SDR while differing by 2× in attack fidelity.
- SCNet-XL is the transient-faithful choice for percussion analysis in this evaluation, rather than the SDR leader.
- The study does not measure perception directly and focuses on MUSDB18-HQ, drums, one checkpoint per architecture, and neutral context length.
6 Conclusion
For rhythm analysis, source separation is usable but not solved by developer metrics alone: each model leaves a systematic signature on drum dynamics and attack. The paper frames separation as an instrument whose transfer function should be measured.
- Music source separation can serve rhythm research, but each model leaves systematic signatures on drum dynamics and attack that carry perceived timing.
Data and code availability
The study provides code, configurations, tables, and figures, with deterministic scripts for regenerating separations, statistics, and figures.
- Code, configurations, per-track tables, and figures are available in the cited repository, alongside MUSDB18-HQ access information.
- Cached separations and analyses can be regenerated deterministically with the listed benchmark and synthesis scripts.