{"id":"ed71c8ab-9fd6-4998-ae71-b6ab9ceda44f","arxiv_id":"2510.16834","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Mamba-based speech enhancer trained with Schrödinger Bridge objectives produces strong denoising and dereverberation in one inference step with a low real-time factor.","lead":"This paper combines Schrödinger Bridge training with a Mamba state-space backbone to enhance noisy, reverberant speech in a single model pass. The authors report one-step results that beat or match several generative and discriminative baselines while running faster, and claim Mamba works better under Schrödinger Bridge than attention or LSTM backbones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled SBCTM baseline: central comparison with a one-step SB method uses a pretrained model not retrained on the same data.","rationale":"The reader's weakest_assumption focuses on the theoretical legitimacy of one-step inference at T=1. That concern is real but partly conceptual: even without an error bound, a well-controlled empirical comparison could still support the headline claim. By contrast, the SBCTM comparison is an explicitly admitted experimental confound in the full text, and it directly targets the 'outperforms strong one-step SB baselines' part of the central claim. It is also concretely testable by retraining the baseline on the same data. The reader's rationale mentions this issue, which is why agreement is partial rather than complete. A controlled SBM-vs-Mamba-base comparison supports the training-paradigm claim, but it does not substitute for a fair one-step SB baseline. Since the reader already returned CONDITIONAL and flagged the SBCTM issue in passing, this stress-test reinforces that verdict without moving it to a different category.","tokens_in":9672,"tokens_out":7103,"duration_ms":60399,"concrete_test":"Retrain SBCTM from its open-source code on exactly the same paired training corpus used for SBM (same 800 h clean speech, same noise set, same RIR set, same SNR range [-10,20], and same STFT settings) and re-evaluate on the four benchmark testsets. If retrained SBCTM reaches or exceeds SBM on DNSMOS, PESQ, and ESTOI in Tables 1-2, the paper's claim of superiority over one-step SB baselines fails; if it remains below SBM on the majority of metrics, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 states: \"SBCTM open-sourced pretrained model and we used it in our experiments. For the other baselines without open-sourced models, ... we re-trained models ... using the same datasets as we used to train SBM.\" Thus the consistency-trajectory one-step SB baseline is not trained on the SBM training distribution. The central claim that SBM \"outperforms strong baselines with 1-step or iterative inference\" depends in part on SBCTM's low scores (e.g., DNS With Reverb PESQ 1.218 and SpeechBERTScore 0.494 vs SBM 1.971 and 0.784; DNS No Reverb OVRL/PESQ where SBCTM is already close or higher). A pretrained checkpoint from a different dataset, noise/RIR simulation, or SNR schedule can explain those gaps without any superiority of SBM. Because SBCTM is one of only two one-step SB variants compared, this is a load-bearing confound for the efficiency claim. The controlled SBM-vs-Mamba-base comparison (same backbone, same data) does not fix it, since Mamba-base is a predictive-mapping model, not a one-step SB method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Schrödinger Bridge Mamba (SBM), a speech-enhancement model that trains a Mamba-based backbone with a Schrödinger Bridge (SB) training objective and performs inference in a single forward pass at timestep T=1. The authors argue that Mamba's Markovian state evolution is naturally compatible with the SB framework, and they evaluate SBM on joint denoising and dereverberation using DNS and VoiceBank-Demand testsets. They report that 1-step SBM outperforms or matches several baselines, including SB-NCSN++, SBCTM, SB-UFOGen, and a predictive-mapping Mamba model (Mamba-base), while achieving the lowest real-time factor in Table 3.","tokens_in":10051,"tokens_out":3577,"duration_ms":31260,"significance":"If the empirical claims hold, the paper makes a useful contribution by replacing the heavy NCSN++ backbone used in most SB-based speech enhancement methods with a much lighter Mamba backbone, thereby enabling one-step inference without sacrificing quality. The controlled comparison against Mamba-base (same backbone, same data, different training paradigm) is the strongest part of the evidence and suggests that the SB training objective itself provides a gain over predictive mapping. The paper also articulates an interesting, though heuristic, conceptual link between SB and selective state-space models. However, the headline claim of 'outperforming strong baselines' is weakened by the uncontrolled SBCTM baseline, the absence of the architecture ablations promised in the abstract, lack of error bars on small metric margins, and an unexamined one-step-at-T=1 approximation. These issues are fixable and do not invalidate the core idea, but they need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The SBCTM baseline is not retrained on the same data. Section 4.3 states that SBCTM's open-sourced pretrained model was used directly, while all other baselines were retrained on the SBM training corpus. Since SBCTM is one of the two one-step SB variants in the comparison, its low scores on DNS With Reverb (e.g., PESQ 1.218 and SpeechBERTScore 0.494 vs SBM 1.971 and 0.784) could be due to mismatched training data, noise/reverberation simulation, or SNR schedule rather than to SBM's superiority. Conversely, SBCTM beats SBM on DNS No Reverb OVRL (3.298 vs 3.292) and PESQ (2.835 vs 2.825), and on VoiceBank PESQ (3.558 vs 3.503). The central comparative claim therefore requires either retraining SBCTM under the same conditions or explicitly restricting the claim to models trained on the same data.","section":"Section 4.3, Tables 1-2"},{"comment":"The abstract promises that 'Ablation studies reveal that the SB paradigm consistently yields improved performance across diverse architectures' and that 'Mamba exhibits a stronger performance under the SB paradigm compared to Multi-Head Self-Attention (MHSA) and Long Short-Term Memory (LSTM) backbones.' The body contains no such ablation table or experiment. Section 5 only compares SBM with Mamba-base, SB-NCSN++, SBCTM, and SB-UFOGen. The claimed MHSA/LSTM comparison is absent, and the 'diverse architectures' claim is unsupported. This missing evidence is load-bearing for the paper's conceptual contribution, not merely a presentation detail.","section":"Abstract vs. Section 5"},{"comment":"No error bars, confidence intervals, or significance tests are reported. Several of the key comparisons involve very small differences, e.g., DNS No Reverb OVRL (SBM 3.292 vs SBCTM 3.298), DNS No Reverb SpeechBERTScore (SBM 0.893 vs Mamba-base 0.887), and VoiceBank PESQ (SBM 3.503 vs SBCTM 3.558, where SBM is worse). Without variance estimates or paired tests, the claim that SBM 'outperforms' these baselines on those metrics is not statistically supported, especially on testsets of only 150–824 utterances. The authors should report per-seed variance, confidence intervals, or at least a paired test on the held-out sets.","section":"Tables 1-2"},{"comment":"The one-step inference design at timestep T=1 is central to the efficiency claim, but it is not analyzed or ablated. Section 3 simply states that 'The inference process is only one step with timestep T=1 representing the start of reverse SDE as in the typical iterative sampling process.' No error bound, distillation argument, or convergence analysis is provided, and no comparison to e.g. T=2, T=5, or T=10 is shown. Without such evidence, it is unclear whether the performance at T=1 reflects a genuine property of SB-trained Mamba or a particular choice of loss and timestep. At minimum, an ablation over inference timesteps would clarify the trade-off and support the 'one-step' claim.","section":"Section 3 and Table 3"}],"minor_comments":[{"comment":"The heading 'EXPERIEMENTAL SETUP' contains a typo; should be 'EXPERIMENTAL SETUP'.","section":"Section 4 heading"},{"comment":"RTF measurement details are insufficient: no GPU model, no batch size, no input duration beyond '10 pieces of 10s audio', and no standard deviation across the 10 pieces. This makes the RTF comparison difficult to reproduce.","section":"Table 3"},{"comment":"Equation (2) presents a linear state-space recurrence, but Mamba's selectivity is usually introduced by making B and C (and sometimes A) input-dependent. The authors mention this informally, but the notation may mislead readers; a brief clarification would help.","section":"Section 2.2"},{"comment":"For the retrained baselines, no details about hyperparameters, training iterations, or validation-based checkpoint selection are given. This makes it hard to assess whether the baselines were tuned to a comparable degree.","section":"Section 4.3"},{"comment":"The statement 'SBM outperforms all baselines in all metrics on the DNS Real Recordings' is supported by Table 2, but the broader claim that SBM 'outperforms strong baselines' is too strong given the exceptions on DNS No Reverb and VoiceBank. The wording should be softened to 'outperforms or matches' where appropriate.","section":"Section 5"}],"recommendation":"major_revision","confidential_remarks":"The abstract-body mismatch regarding the MHSA/LSTM ablations is a serious inconsistency that should be caught in the review process. If the missing ablations are simply an oversight in the submitted version, adding them may be straightforward. The uncontrolled SBCTM comparison is the most significant technical weakness and should be fixed or the claims must be scoped accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper does one concrete thing: it takes the Schrödinger Bridge training objective and applies it to a Mamba backbone for speech enhancement, achieving 1-step inference with a small 3.93M model and RTF 0.0048. The controlled comparison against a same-backbone predictive-mapping model (Mamba-base) is the strongest part: it suggests SB training helps, and that result is worth having.\n\nThe novelty is modest but real — no one else has published this exact pairing. The efficiency numbers are attractive, though measured on a single GPU with 10 samples and no variance reported.\n\nThe soft spots are substantial. First, the abstract promises ablations across Mamba, MHSA, and LSTM, but the body only reports Mamba-base vs SBM. Those other architectures are absent. Second, the SBCTM baseline was not retrained on the same data; it uses an open-sourced checkpoint. That matters because on DNS No Reverb and VoiceBank, SBCTM actually beats SBM on some metrics (e.g., PESQ on VoiceBank: 3.558 vs 3.503), so the sweeping claim that SBM 'outperforms strong baselines' is not uniformly supported. The stress-test note about SBCTM is accurate, and it weakens the efficiency/quality argument. Third, the one-step inference at T=1 is justified heuristically; there is no error bound or convergence analysis. It works in practice, but the paper doesn't tell you why it should.\n\nThere are also small things: no error bars anywhere, no code release, no training hyperparameters. The compatibility argument between SB and Mamba is qualitative — a 'plausible conjecture' — and doesn't rise to a theory, though the paper doesn't claim more than that.\n\nOverall, the central intuition is plausible and the controlled Mamba-base result is a genuine contribution. But the paper needs major revision: run the promised ablations, retrain SBCTM (or at least document its training data and discuss the mismatch), report variance, and soften the superiority claims. As it stands, it's a solid workshop-level or short-paper result that overclaims in a few places.\n\nI'd send it to peer review — the combination is timely and the core experiment is worth scrutinizing — but I would not cite it for its headline numbers until the baseline issue is fixed.\n\nBest.","headline":"Useful SB+Mamba combination for one-step speech enhancement, but the published claims run ahead of the evidence.","tokens_in":10442,"tokens_out":2631,"would_cite":false,"duration_ms":23907,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that training a Mamba backbone with a Schrödinger Bridge objective collapses the optimal transport path from degraded to clean speech into a single inference step, outperforming generative baselines that need 10-50 steps wh","keywords":["Schrödinger Bridge","Mamba","speech enhancement","one-step inference","generative speech enhancement","denoising","dereverberation","state-space models"],"falsifier":"Run the trained SBM in iterative mode (e.g., taking multiple small reverse steps with the same network) and compare output quality to the one-step output; if iterative inference clearly outperforms one-step, the one-step claim is falsified. Alternatively, evaluate one-step SBM on noise conditions with SNRs far outside the training range; a sharp quality collapse would indicate the model approximates a mapping rather than the true transport.","tokens_in":9557,"feed_emoji":"🔊","tokens_out":6691,"duration_ms":52466,"temperature":0.7,"pith_summary":"The paper proposes Schrödinger Bridge Mamba (SBM), a speech enhancement model that trains a Mamba state-space backbone under a Schrödinger Bridge objective. The central claim is that Mamba's Markovian dynamics and the Schrödinger Bridge's stochastic process align, so a single forward pass at the start of the reverse process maps degraded speech to clean speech without iterative refinement. In a joint denoising and dereverberation task spanning four test sets, SBM with one inference step outperforms or matches Schrödinger Bridge models that require 10-50 steps, as well as one-step Schrödinger Bridge variants and a predictive-mapping Mamba baseline, while achieving the smallest real-time factor. If this holds, generative speech enhancement can be high-quality and real-time even on lightweight devices.","feed_headline":"One-step Schrödinger Bridge Mamba outpaces iterative speech models","feed_subtitle":"Schrödinger Bridge training lets a Mamba state-space model clean speech in one forward pass at the fastest runtime.","key_machinery":"The key machinery is the SBM training-inference scheme: the Schrödinger Bridge objective trains a Mamba backbone using data-prediction losses on interpolated states, with a timestep embedding added to the Mamba blocks; inference then runs the backbone once at timestep T=1. The compatibility between SB and Mamba is the Markov property shared by the SB reverse process and Mamba's state recurrence h_t = A h_{t-1} + B u_t, which lets the model act as an implicit one-step optimal transport map.","core_discovery":"SBM's discovery is that the Schrödinger Bridge training paradigm, previously tied to heavy score-network backbones and iterative sampling, can be paired with a selective state-space backbone so that the optimal transport between degraded and clean speech is collapsed into a one-step state transition. The paper argues this works because Mamba's hidden-state evolution is Markovian, matching the SB stochastic process, and its selective mechanism can dynamically parameterize the optimal control along the path. Empirically, the one-step SBM outperforms baselines on perceptual, semantic, and speaker-similarity metrics, with a real-time factor of 0.0048 and only 3.93M parameters.","pith_inferences":["The one-step jump from T=1 is an empirical shortcut; the paper gives no error bound, so the claim that a single forward pass faithfully approximates the Schrödinger Bridge transport rests entirely on experimental evidence.","The Markov-property compatibility argument is suggestive but not a formal equivalence proof; a careful study of how Mamba's discretized SSM relates to the SB SDE could turn the architectural alignment into a provable guarantee.","If SBM is extended to other tasks, the same training recipe (SB objective + timestep embedding + one-step inference) is a testable baseline for image, video, and cross-modal generation.","Direct ablations that isolate the architecture from the training paradigm (e.g., same Mamba backbone trained with predictive mapping vs. SB, and same SB objective with different backbones) would further support the claimed synergy."],"forward_implications":["One-step inference makes generative speech enhancement practical for real-time and edge applications, with a real-time factor below 0.005.","Switching from predictive mapping to SB training improves the same Mamba backbone on all test sets, with no increase in inference cost.","The same SB+Mamba concept could be applied to other generative restoration tasks such as super-resolution and inpainting.","Because SBM inherits Mamba's streaming and linear-complexity properties, it can be further optimized for streaming devices.","The method closes the performance gap between one-step generative enhancement and slower iterative models."],"fun_headline_variants":["One-step SBM cleans speech in a single pass","Mamba + Schrödinger Bridge: speech cleanup in one pass","Single forward pass: Schrödinger Bridge Mamba denoises speech","SBM: real-time speech enhancement with a single step","Schrödinger Bridge Mamba: one-step speech enhancement"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire efficiency and quality claim rests on the assumption that one forward pass of the Mamba backbone at timestep T=1 faithfully reproduces the full multi-step Schrödinger Bridge reverse transport, yet the paper offers no proof, error bound, or distillation argument for that jump.","fun_headline_variants_meta":{"raw":{"variants":["One-step SBM cleans speech in a single pass","Mamba + Schrödinger Bridge: speech cleanup in one pass","Single forward pass: Schrödinger Bridge Mamba denoises speech","SBM: real-time speech enhancement with a single step","Schrödinger Bridge Mamba: one-step speech enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002238,"raw_usage":{"total_tokens":8453,"prompt_tokens":672,"completion_tokens":7781,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":416,"completion_tokens_details":{"reasoning_tokens":7699}},"tokens_in":416,"tokens_out":7781,"duration_ms":44088,"temperature":1.0,"reasoning_tokens":7699,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:08:59.608830+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained SBM in iterative mode (e.g., taking multiple small reverse steps with the same network) and compare output quality to the one-step output; if iterative inference clearly outperforms one-step, the one-step claim is falsified. Alternatively, evaluate one-step SBM on noise conditions with SNRs far outside the training range; a sharp quality collapse would indicate the model approximates a mapping rather than the true transport.","supporting_citations":[],"review_version":1}