SE-MSB: End-to-End Unpaired Speech Enhancement using Mamba Schrödinger Bridges
Andreas Bagge, Andreas Nymand, Michael Riis Andersen, Bjørn Sand Jensen
eess.AS · cs.AI · cs.SD
Quick verdict
Proposes first fully unpaired raw-waveform SE via DSB+Mamba; efficiency and few-step claims well supported, but 'on par or better' overstates vs BUDDy.
Summary
The paper presents SE-MSB, a speech-enhancement framework trained fully unpaired — without clean/degraded pairs — by learning a Diffusion Schrödinger Bridge between clean and degraded speech distributions, implemented as a Mamba2-based diffusion network operating directly on raw 16 kHz waveforms (1D-convolutional frontend, ten Mamba blocks, ~45.6M parameters). Training follows DSB Matching with a pre-training phase on random couplings and a fine-tuning phase on simulated couplings, 350k steps on one RTX 4090. Across dereverberation, denoising, declipping, and compounded degradations on VCTK, LibriSpeech, and CHiME-6, SE-MSB outperforms unpaired baselines GFB and WPE on most metrics and decisively beats BUDDy once degradations violate its convolutional assumption (SISDRi 4.95 vs 0.73 on reverb+noise). Its headline contribution is efficiency: roughly 80x fewer FLOPs and ~100x lower measured latency than BUDDy, with quality nearly unchanged between 10 and 50 sampling steps. However, BUDDy(100) still leads on four of five VCTK dereverberation metrics, and on LibriSpeech SE-MSB worsens WER (0.11 vs 0.03 for unprocessed audio) and speaker similarity, straining the generalization claim. Supervised models (SGMSE, UNIVERSE++, Sepformer) remain ahead on most metrics. Limitations include no significance testing, no subjective listening test, and minor protocol inconsistencies.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 2
● low 4
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- SE-MSB is the first fully unpaired, end-to-end speech enhancement model operating on raw waveforms, combining DSB Matching with a Mamba2 diffusion backbone (Table 1).
- SE-MSB matches or exceeds unpaired baselines GFB and WPE on most metrics and is roughly 80x cheaper in FLOPs and ~100x faster in measured latency than BUDDy (Tables 2-3).
- Output quality is nearly invariant to the number of sampling steps between 10 and 50, enabling few-step inference without performance loss (Table 2, Figure 4).
- The DSB formulation transfers across degradation types: on compounded reverb+noise, one unpaired model reaches SISDRi 4.95 vs BUDDy's 0.73 (Table 4).
- The gap to paired supervised methods is negligible for some metrics, making unpaired adaptation practical for in-the-wild acoustic environments.
Strengths
Genuine gap filled
Table 1 and §3 position SE-MSB as the first fully unpaired, end-to-end raw-waveform SE model; prior unpaired diffusion methods give up at least one property (GFB uses an internal STFT encoder/ISTFT decoder, BUDDy works on spectrograms), and the DSB-Matching + Mamba2 combination is new to the waveform domain.
Efficiency evidence is concrete and multi-dimensional
Table 3: SE-MSB(10) needs 2.39 TFLOPs and 214 ms vs BUDDy(100) at 192 TFLOPs and 22,135 ms (~80x / ~103x), at comparable parameter counts (45.6M vs 27.7M); the Mamba2 selective-scan FLOP accounting is actually derived in Appendix D rather than asserted.
Few-step robustness demonstrated, not asserted
Compounded-degradation result is the strongest scientific finding
Broad, mostly fair evaluation protocol
Weaknesses
Abstract overclaims relative to Table 2
Generalization claim undercut by own OOD numbers
Statistical reporting thin for the number of comparisons
Internal inconsistencies and undefined notation
Efficiency framing is selective
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SE-MSB (a NeurIPS 2026 submission) targets speech enhancement — dereverberation, denoising, declipping — trained fully unpaired: clean (VCTK) and degraded distributions are sampled independently, so the model never observes a (clean, degraded) pair (§4). The method marries Diffusion Schrödinger Bridge Matching (refs 16, 17) with a Mamba2 diffusion backbone applied to raw waveforms. Central claims: (i) first fully unpaired end-to-end raw-waveform SE model (Table 1); (ii) on par or better than paired and unpaired state of the art; (iii) orders-of-magnitude faster inference; (iv) task flexibility across compounded degradations. Claim (ii), as phrased in the abstract, is stronger than what Table 2 shows — a recurring theme of this review.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SE-MSB and what problem does it solve?
SE-MSB is a speech enhancement model that cleans noisy, reverberant, or clipped speech without ever training on clean/degraded audio pairs. It learns a diffusion Schrödinger bridge — a stochastic transport between the clean and degraded speech distributions — using an efficient Mamba (state-space) network that operates directly on raw 16 kHz waveforms, so it can adapt to real-world degradations that cannot be simulated for supervised training.
How much faster is SE-MSB than BUDDy?
In the paper's measurements on 2.048 s of audio, SE-MSB with 10 sampling steps needs 2.39 TFLOPs and 214 ms, versus 192 TFLOPs and 22,135 ms for BUDDy with 100 optimization steps — roughly 80x fewer FLOPs and ~100x lower latency. Against the other unpaired baseline GFB the speedup is only about 1.5x, so the 'orders of magnitude' claim is specific to BUDDy.
Does SE-MSB beat supervised speech enhancement models?
Not overall. Supervised baselines such as SGMSE, UNIVERSE++, and Sepformer still lead on most metrics. SE-MSB narrows the gap on some (e.g., FAD 0.31 vs UNIVERSE++ 0.27 on VCTK dereverberation) and posts the best DNSMOS of all tested models on real CHiME-6 recordings, where the supervised models degrade below even unprocessed audio.
Is the code and data for SE-MSB available?
The paper links an anonymous code repository (anonymous.4open.science/r/Latent-DSB) with audio examples. All datasets used — VCTK, public RIR collections, WHAM! noise, LibriSpeech — are publicly available, and full training hyperparameters are given in the appendix, though there is a batch-size inconsistency between the main text (18) and the appendix (8, effective 16).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are available only as captions in the extracted text; the compute-vs-quality trade-off (Figure 2) and full step-count curves (Figure 4) cannot be independently verified.
- The footnote symbol on the Paired SE-MSB rows of Table 2 is undefined in the extracted text.
- Table extraction interleaves row-block labels (VCTK/Libri/Reverb+Noise); the block mapping was inferred from layout and could shift a metric by one row in rare cases.
- Appendix D as extracted covers computational efficiency only; the referenced CHiME-6 equivalent plot is not present.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.