{"id":"ab118810-60af-4b1e-bdc5-45f73aa06c96","arxiv_id":"2412.19099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-branch, band-split speech enhancement network with Mamba achieves competitive quality at a fraction of the computational cost of transformer-based models.","lead":"This paper introduces a speech enhancement network that cleans noisy speech by processing its loudness and phase information in two separate branches, then recombining them. It uses a Mamba state-space model and splits the frequency range into bands to cut computation, reporting roughly 8x lower cost than the average baseline and 25x lower than a transformer model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '8.3x complexity reduction' is computed over a baseline set that includes DBTNet, yet BSDB-Net is 0.26 PESQ worse than DBTNet; restricting to comparable-performance baselines shrinks the advantage to about 2-3x, so the 'superior performance' claim is unsupported.","rationale":"The reader's verdict of CONDITIONAL is appropriate, but my most load-bearing concern differs from the reader's chosen weakest assumption. The reader flagged the missing band-split ablation; that concern is valid but secondary. The more direct threat to the central claim is the internal inconsistency between the abstract's 'superior performance' and the paper's own admission that 'our model is only slightly behind DBTNet' (Model Complexity Comparison section). The 8.3x figure is an average over a mixed set of baselines, many of which are both heavier and weaker; the 25x figure is obtained specifically against DBTNet, which outperforms BSDB-Net by 0.26 PESQ. Thus the headline complexity reduction is real only when measured against the full baseline list, not when the comparison is restricted to models of comparable or better quality. Table 1 omits DBTNet, so the SOTA claim is not representative of the strongest available reference. Additional data-quality issues, such as the Set-A PESQ row in Table 1 (2.43, 2.92, 2.26 with average 2.87, where 2.26 is likely a typo for 3.26), further reduce confidence in the reported numbers but do not by themselves overturn the architecture's value. The paper's positive contributions are that the Mamba-based dual-branch design is clearly specified, the complexity figures are plausibly reproducible, and the band-split strategy is a sensible way to reduce frequency modeling cost. Provided the authors clarify the comparison set, correct the table typos, and reword 'superior performance' to 'comparable performance' where appropriate, the paper remains a reasonable conditional acceptance.","tokens_in":14046,"tokens_out":5882,"duration_ms":51475,"concrete_test":"Recompute the average MACs ratio in Table 2 using only baselines whose PESQ is within 0.05 of BSDB-Net's 2.92 (e.g., CTSNet 2.79, GaGNet 2.85, DDAEC 2.79) and report the largest PESQ gap against any baseline. If the average ratio drops below 4x or the gap to DBTNet exceeds 0.1 PESQ, the abstract must be reworded to 'comparable performance' instead of 'superior performance'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the performance-complexity trade-off: 'average 8.3 times reduction in computational complexity while maintaining superior performance' (Abstract). Table 2 lists DBTNet at PESQ 3.18 with 42.64 G/s and BSDB-Net (128-6) at PESQ 2.92 with 1.68 G/s. The 8.3x figure is the average over all baselines in Table 2, including DBTNet, DDAEC, and FullSubNet, which have very high MACs; the 25x figure is specifically DBTNet/1.68 = 25.4. On the strongest baseline, BSDB-Net is 0.26 PESQ worse. Table 1, which claims SOTA on WSJ0-SI84, omits DBTNet entirely, so the SOTA claim is only against a weaker subset. If one computes the complexity ratio only against models with PESQ at least 2.85 (GaGNet 2.81 G/s, CTSNet 5.48 G/s), the reduction is about 1.7-3.3x, not 8.3x. The abstract's 'superior performance' is therefore not robust to the choice of comparison set, and the headline complexity ratio conflates average compute over heavy baselines with a genuine Pareto improvement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BSDB-Net, a causal monaural speech enhancement model that decouples magnitude and complex-spectrum enhancement in two parallel branches, compresses the frequency axis with a band-split module, and uses Mamba-based blocks for time and frequency sequence modeling. The method is evaluated on a WSJ0-SI84+DNS-Challenge corpus and on VoiceBank+Demand. The authors report state-of-the-art or competitive PESQ, ESTOI, and SI-SDR results and claim an average 8.3x reduction in computational complexity relative to baselines, with the best 128-6 configuration at 1.68 G/s and PESQ 2.92.","tokens_in":14364,"tokens_out":9033,"duration_ms":76917,"significance":"If the reported numbers are taken at face value after correcting the Table 1 typo, the paper makes a useful contribution to efficient speech enhancement: it demonstrates that a causal dual-branch architecture with band-split frequency compression and Mamba sequence modeling can reach the performance level of much heavier T-F-domain models on WSJ0-SI84 and VoiceBank+Demand. The ablations of the dual-branch structure and the Mamba-versus-LSTM/Transformer replacement are informative, and the complexity table using a fixed one-second input is a useful step toward fair MAC comparisons. The central efficiency claim, however, is currently overstated relative to the evidence in Tables 1 and 2, so the significance depends on how the authors choose to scope the claim.","major_comments":[{"comment":"The BSDB(ours) PESQ values for Set-A are 2.43, 2.92, and 2.26 at -5, 0, and 5 dB, with an average reported as 2.87; arithmetic requires the 5 dB value to be approximately 3.26, not 2.26. Because the row average and the subsequent state-of-the-art claim depend on this value, please correct the cell and re-verify all metrics in the table.","section":"Table 1, Set-A, BSDB row"},{"comment":"The claim of an average 8.3x complexity reduction while maintaining superior performance is not supported when the comparison is restricted to baselines with comparable PESQ. The strongest baseline in Table 2, DBTNet, reaches PESQ 3.18 at 42.64 G/s whereas BSDB-Net (128-6) reaches PESQ 2.92 at 1.68 G/s, a 0.26 PESQ loss; against GaGNet (2.85 PESQ, 2.81 G/s) and CTSNet (2.79 PESQ, 5.48 G/s) the reduction is roughly 1.7-3.3x, not 8.3x. Please report the complexity ratio relative to comparable-performance baselines and align the abstract wording with the actual comparison set, since Table 1 omits DBTNet entirely.","section":"Abstract and Table 2"},{"comment":"The ablation section covers the dual-branch structure, the number of Mamba layers and hidden dimensions, and Mamba versus LSTM/Transformer, but it does not test the band-split compression itself. Because band-splitting is one of the two mechanisms behind the claimed complexity reduction, an experiment that varies K or bypasses band-splitting is needed to show that the frequency compression does not discard information necessary for phase recovery; without it, the performance-complexity trade-off cannot be attributed to the proposed design.","section":"Ablation Study and Eqs. (2)-(4)"},{"comment":"The text states that the model compresses computational complexity 'to the level of Mb/s', but Table 2 reports all BSDB-Net variants between 0.88 G/s and 4.26 G/s. Please correct the units, and also make the complexity claim numerically consistent with the abstract's '8.3 times' by reporting the exact average and the baseline set over which it is computed.","section":"The Number of Layers and Hidden Layers of Mamba-Block and Table 2"}],"minor_comments":[{"comment":"The noisy PESQ value at 0 dB is written as '2,04' and should be '2.04'.","section":"Table 1, Set-B, noisy row"},{"comment":"The heading 'Modle' should be 'Model', and the DBTNet row uses '%' in the Cau. column instead of a check mark; the text should explicitly state that DBTNet is non-causal.","section":"Table 2"},{"comment":"The text refers to the 'BSDS framework' where 'BSDB' is intended.","section":"Proposed Architecture"},{"comment":"The reference 'Chan, J. N. L. Q., W. 2016' is malformed and should give the full author list.","section":"References"},{"comment":"The SSM recurrence is stated without the discretization equations; please include them or cite the Mamba paper at that point.","section":"Eqs. (17)-(18)"},{"comment":"Please report standard deviations or significance tests for the main metrics, since several reported differences are small and the state-of-the-art claim rests on them; also standardize the names 'CTS-Net'/'CSTnet' and 'S4NDUNet'/'S4DSE'.","section":"Tables 1 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a speech/audio machine-learning venue, and the core architecture is reasonable. The main gap is that the efficiency claim in the abstract is broader than what Tables 1-2 support, and the missing band-split ablation makes it hard to attribute the trade-off to the proposed contribution. If the authors correct the Table 1 typo, add the band-split ablation, and rescope the complexity claim to comparable-performance baselines, the paper would be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead BSDB-Net. The short version: the exact combination of band-split, dual-branch magnitude/complex modeling, and Mamba is new, and it delivers a causal model with 1.68 G/s at PESQ 2.92 on WSJ0-SI84, which is a real efficiency point. The ablations are useful: dual-branch helps, Mamba beats LSTM/Transformer at a fraction of the MACs, and the VoiceBank+Demand results are competitive with much heavier models. The paper does what a good systems paper should: it shows a working cheaper alternative, not a theoretical breakthrough.\n\nBut the headline numbers do not survive close reading. The abstract says “8.3 times reduction … while maintaining superior performance.” The paper’s own complexity section says the model is “only slightly behind DBTNet.” That is a contradiction. DBTNet scores 3.18 PESQ at 42.64 G/s; BSDB-Net (128-6) scores 2.92 at 1.68 G/s. So on the strongest baseline, you lose 0.26 PESQ. The 8.3x is an average over very heavy baselines, including DBTNet, DDAEC, and FullSubNet. Restrict the comparison to models within 0.1 PESQ of yours and the complexity advantage shrinks to maybe 2-3x. That is still a real advantage for a causal model, but it is not what the abstract claims. Table 1 also omits DBTNet entirely, so the “SOTA on WSJ0” claim is against a weaker subset.\n\nOther issues are smaller but real. Table 1 has a likely typo: Set-A BSDB PESQ at 5 dB is 2.26, which breaks the row average of 2.87. No error bars or significance tests anywhere. No code or checkpoint released, which makes the complexity numbers hard to verify. And there is no ablation that removes the band-split compression, so we do not know if frequency banding is discarding phase-relevant details. That last point is a genuine gap given the dual-branch design depends on phase recovery.\n\nWho is this for? Researchers working on efficient causal speech enhancement, and people comparing Mamba-based sequence models in audio. It is a legitimate data point, but it needs a careful reader. I would not cite it as-is because the central trade-off claim is misleading, though the architecture itself is worth a second look after revision.\n\nRecommendation: send it to peer review, but with a clear ask. The authors need to restate the complexity comparison using matched-performance baselines, add error bars, fix the table typo, and ideally release code. If they do that, this becomes a solid systems contribution. Right now it is a good idea wrapped in an overclaimed headline.\n\nBest,\n[You]","headline":"A genuinely cheap causal SE architecture worth knowing about, but the abstract's complexity/perfomance claim is overstated and needs fixing before this is citable.","tokens_in":14869,"tokens_out":1968,"would_cite":false,"duration_ms":20672,"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":"This paper claims a dual-branch, band-split network using Mamba sequence modeling can cut computational cost of monaural speech enhancement by an average factor of 8.3 while matching or exceeding state-of-the-art quality.","keywords":["speech enhancement","monaural","Mamba","selective state space models","band-split","dual-branch","amplitude-phase decoupling","complex spectral mapping"],"falsifier":"Replace the band-split in BSDB-Net with the full FFT resolution (K equal to the number of frequency bins) on WSJ0-SI84 Set-B while keeping everything else fixed; if PESQ, ESTOI, or SI-SDR improve by a large margin, the compression is discarding information the enhancement depends on. Independently recomputing the Table 2 MACs on one-second inputs with standard profiling would confirm the 8.3x and 25x reduction ratios.","tokens_in":13877,"feed_emoji":"🎙️","tokens_out":4764,"duration_ms":195598,"temperature":0.7,"pith_summary":"This paper aims to solve two problems at once in monaural speech enhancement: the compensation effect caused by jointly estimating amplitude and phase from complex spectra, and the high computational cost of modern enhancement models. It proposes a dual-branch network in which one branch enhances the magnitude spectrum while the other estimates the complex spectrum, with an interaction module that lets each branch suppress or supply what the other lacks. To cut cost, it groups frequency bins into non-overlapping bands and models the resulting sequences with Mamba, a selective state-space model that runs in linear time. On WSJ0-SI84 and VoiceBank+Demand, the authors report matching or better quality than earlier methods while cutting average complexity by 8.3 times relative to baselines and 25 times relative to transformer-based models. The payoff, if right, is a path to high-quality speech enhancement that fits on devices with tight compute budgets.","feed_headline":"Dual-branch Mamba slashes speech-enhancement compute 8.3x","feed_subtitle":"Amplitude and phase get separate branches while band-splitting and Mamba keep quality high at low cost.","key_machinery":"Three components carry the argument. The Band-Split module compresses the F frequency bins into K non-overlapping bands, each projected to dimension N, producing a compact tensor for sequence modeling, and the Mask-Decoder merges bands back into a full-spectrum mask. The dual-branch structure consists of a Magnitude Enhancement Network and a Complex Spectral Enhancement Network running in parallel, with an Interaction module that concatenates the two streams, creates a sigmoid mask, and lets one branch repair missing components of the other. The Mamba-Block performs bidirectional frequency modeling and unidirectional causal time modeling using selective state-space equations, giving linear complexity in both axes.","core_discovery":"The central claim is that decoupling magnitude and phase in parallel branches, rather than mapping a single complex spectrum to another one, removes the compensation effect in which amplitude is degraded to repair phase. The magnitude branch applies a mask to roughly suppress noise; the complex branch estimates real and imaginary components; and the two streams are combined after an interaction step. Band-splitting compresses the frequency dimension before sequence modeling, and Mamba-based blocks model time and frequency with linear complexity while keeping the model causal. With its chosen configuration, the paper reports quality scores above all compared baselines on WSJ0-SI84 and VoiceBank+Demand while running at 1.68 G/s MACs, and states that this is state-of-the-art across all metrics.","pith_inferences":["The published 8.3x average is computed over the chosen baselines and excludes the non-causal transformer model DBTNet; extending the comparison to that model would likely lower the headline ratio.","The band-split count K is a tunable knob the paper does not explore in depth; finer bands would probably help difficult, phase-heavy conditions at higher cost.","The interaction module's specific contribution is not isolated in the ablations, so its role in suppressing noise versus restoring signal remains open.","The dual-branch decoupling idea may transfer directly to other speech tasks such as dereverberation and target extraction, which the paper lists as future work."],"forward_implications":["Speech enhancement can be pushed toward real-time, low-power deployment because the model operates at megabyte-scale per second rather than gigabyte-scale per second.","Causal, streaming enhancement is compatible with this design because the time-dimension Mamba is unidirectional.","Scaling behavior favors depth over width, so a smaller hidden dimension with more stacked Mamba blocks can be chosen to balance quality and cost.","Replacing quadratic-attention transformers with Mamba in other time-frequency sequence tasks may produce large complexity savings with only a small quality difference."],"supporting_citations":[{"why":"Supplies the band-split and band-merge design that compresses the frequency axis before modeling.","marker":"(Yu and Luo 2023)"},{"why":"Provides the selective state-space model, Mamba, used as the linear-complexity sequence modeling core.","marker":"(Gu and Dao 2023)"},{"why":"Establishes the compensation effect between magnitude and phase that the dual-branch design targets.","marker":"(Wang, Wichern, and Le Roux 2021)"},{"why":"Introduces CTSNet, a two-stage decoupling baseline, and supplies the WSJ0-SI84 dataset synthesis used in experiments.","marker":"(Li et al. 2021a)"},{"why":"Provides GaGNet, a parallel amplitude-phase decoupling baseline with strong results that BSDB-Net compares against.","marker":"(Li et al. 2022b)"},{"why":"Gives DBTNet, the non-causal transformer-based model used as a high-complexity performance reference.","marker":"(Yu et al. 2022)"},{"why":"Provides GCRN, a complex spectral mapping baseline representing the single-branch approach the paper argues against.","marker":"(Tan and Wang 2019)"}],"fun_headline_variants":["Dual-branch Mamba cuts speech-enhancement compute 8.3x","Decoupled amplitude and phase improve speech enhancement","Band-split Mamba shrinks speech-enhancement compute 8.3x","Speech enhancement: separate phase and amplitude improves quality","Mamba-based speech enhancer: 8.3x less compute, state-of-the-art"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that grouping frequency bins into K fixed non-overlapping bands loses no information the enhancement needs, since the paper tests band-splitting but not a no-band-split version.","fun_headline_variants_meta":{"raw":{"variants":["Dual-branch Mamba cuts speech-enhancement compute 8.3x","Decoupled amplitude and phase improve speech enhancement","Band-split Mamba shrinks speech-enhancement compute 8.3x","Speech enhancement: separate phase and amplitude improves quality","Mamba-based speech enhancer: 8.3x less compute, state-of-the-art"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4352,"prompt_tokens":920,"completion_tokens":3432,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":3334}},"tokens_in":536,"tokens_out":3432,"duration_ms":27249,"temperature":1.0,"reasoning_tokens":3334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:57:24.796515+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the band-split in BSDB-Net with the full FFT resolution (K equal to the number of frequency bins) on WSJ0-SI84 Set-B while keeping everything else fixed; if PESQ, ESTOI, or SI-SDR improve by a large margin, the compression is discarding information the enhancement depends on. Independently recomputing the Table 2 MACs on one-second inputs with standard profiling would confirm the 8.3x and 25x reduction ratios.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides GCRN, a complex spectral mapping baseline representing the single-branch approach the paper argues against."}],"review_version":1}