{"id":"6c38752a-77b7-477e-b676-d07279c83797","arxiv_id":"2505.16911","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Transformer-Mamba model that adds a learned correction signal to degraded speech beats adapted active-noise-control baselines on denoising, dereverberation, and declipping in simulation.","lead":"Active Speech Enhancement is proposed as a new paradigm: rather than only canceling noise as in active noise cancellation, a model computes a correction signal that is added to degraded speech to restore a clean target. The paper presents a Transformer-Mamba model that outperforms three adapted active-noise-control baselines on denoising, dereverberation, and declipping in a simulated acoustic room.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Runtime section claims real-time viability with 31 ms future-frame prediction, but the stated geometry allows only ~4.3 ms of processing delay, so the reported gains may reflect non-causal lookahead rather than active shaping.","rationale":"The reader correctly identified physical realizability as the weakest assumption. I refine this to a specific, checkable failure mode: the paper's own runtime analysis is internally inconsistent. All the empirical comparisons in Tables 1-3 are offline; the runtime section is the only place where causality is discussed, and there the proposed 31.25 ms lookahead exceeds the available 4.3 ms budget by roughly a factor of seven. Because a feedforward ANC/ASE controller in a 3x4x2 m room with the stated microphone and loudspeaker coordinates cannot observe 31 ms into the future, the main results cannot honestly be described as real-time active enhancement. This matters most for dereverberation and declipping, where the target c(n)=s(n) must be recovered from a signal that has been filtered by the primary path P(z); inverting P(z) through the secondary path S(z) generally requires non-causal filtering, and the allowed lookahead is precisely what could provide it. The proposed strict-causality re-evaluation would settle whether the reported gains are inherent to the ASE principle or an artifact of off-line processing. I am not claiming the authors are deceptive; the limitation section already concedes baseline adaptation and per-task training. The narrow offline claim may well survive, so I do not recommend rejection. However, the real-time and active framing should be conditional on demonstrating causal performance or on explicitly relabeling the results as offline-only.","tokens_in":17019,"tokens_out":15695,"duration_ms":138113,"concrete_test":"Retrain or evaluate ASE-TM with a strict causal mask: for each output frame at time n, allow only input frames up to n + D, with D corresponding to at most 4.3 ms (about 69 samples at 16 kHz), or zero future frames, while keeping all other hyperparameters and the task-generation pipeline identical. Recompute Tables 1-3 for ASE-TM, THF-FxLMS, DeepANC, and ARN under this constraint. If causal ASE-TM's PESQ remains near 2.96 and still exceeds ARN, the real-time concern does not land; if it drops toward or below the ANC baselines, the reported gains are largely an artifact of non-causal lookahead.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing vulnerability is the causal feasibility of the active system. Section 6.4 gives the feedforward causality budget as T_ASE-TM < T_p - T_s ≈ 4.3 ms, then says the model predicts 500 future frames (0.03125 s) to accommodate inference latency and calls this real-time. A 31.25 ms lookahead is not available in the described geometry: with the reference and modification microphones 2 m apart and the secondary loudspeaker 0.5 m from the modification microphone, the controller cannot know 31 ms of future audio before the primary disturbance reaches the modification microphone. The main tables (Tables 1-3) do not state a causal constraint, so the reported PESQ and NMSE gains may be obtained with full-utterance, non-causal processing. Since THF-FxLMS is causal by construction and ARN/DeepANC are low-latency ANC methods, the comparison may conflate the effect of future context with the contribution of active speech shaping. The dereverberation and declipping results are especially suspect, because matching the target c(n)=s(n) from an observed P(z)-filtered signal requires inverting P(z) through S(z); such an inversion generally needs non-causal filtering. If the model's advantage disappears under a strict causality constraint, the central active claim is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Active Speech Enhancement (ASE), a paradigm in which a neural controller generates a correction signal rendered by a loudspeaker and added acoustically to the degraded speech at an error (modification) microphone. The proposed ASE-TM model is a Transformer-Mamba network trained with a composite loss including a PESQ-predicting adversarial discriminator. Experiments in a simulated room (image-method RIRs, SEF loudspeaker nonlinearity) on VoiceBank-DEMAND denoising, dereverberation, and declipping compare ASE-TM with THF-FxLMS, DeepANC, and ARN. The paper reports superior PESQ and secondary metrics in all three tasks, plus a runtime analysis claiming real-time feasibility via future-frame prediction.","tokens_in":17323,"tokens_out":7759,"duration_ms":61223,"significance":"If the claims were fully supported, the paper would offer a new and potentially useful extension of ANC to speech enhancement, with a clear formalization and a strong architecture. The authors provide a detailed simulation pipeline, an ablation study, and a robustness analysis across T60 and loudspeaker nonlinearity. However, the current evidence does not establish that the active paradigm is responsible for the gains: the baselines are all ANC methods, the headline metric is directly optimized, and the real-time claim is contradicted by the stated geometry and lookahead budget. The dereverberation and declipping targets are physically inconsistent with the measured signal. These problems are substantial but fixable with new experiments and analysis.","major_comments":[{"comment":"The claim in the abstract that ASE-TM \"outperforms existing baselines\" is not supported by the chosen comparators. All three baselines (THF-FxLMS, DeepANC, ARN) are ANC methods adapted to the ASE framework; no standard speech enhancement model is evaluated. In particular, ASE-TM is built on the SEmamba architecture (§4.1) and uses the MP-SENet/SEmamba loss framework, yet neither SEmamba nor any passive SE model (e.g., CMGAN or MetricGAN+) is run on the same simulated degraded signal d(n). A direct comparison with SEmamba applied to d(n) would isolate whether the active acoustic addition contributes anything beyond the backbone and loss. As it stands, Tables 1–3 support only the narrow claim that ASE-TM beats adapted ANC baselines.","section":"§5.4, Tables 1–3"},{"comment":"The causality claim is internally inconsistent. The paper states the feedforward causality condition is T_ASE-TM < Tp − Ts ≈ 0.0043 s, then says the model predicts 500 future frames (0.03125 s) to accommodate inference latency while \"remaining within real-time limits.\" The lookahead is roughly 7× larger than the stated processing budget, so the implementation is non-causal in the described geometry. Because Tables 1–3 do not state whether the reported results are causal or full-utterance/non-causal, the gains may reflect future context rather than active speech shaping. The comparison with causal THF-FxLMS and low-latency ARN is therefore confounded; the experiments should be rerun with lookahead limited to the available budget, or the non-causal results should be explicitly labeled as an offline upper bound.","section":"§6.4"},{"comment":"The target definition for dereverberation and declipping is physically problematic. For additive noise, c(n) = s(n)*P(z), which is consistent with eh(n) being measured after the primary path. For dereverberation and declipping, however, c(n) is defined as the original anechoic, unclipped clean speech \"prior to any acoustic path effects.\" Since eh(n) = d(n) + a(n) is measured at the modification microphone after P(z), driving eh(n) toward c(n) requires the secondary path S(z) to invert, at least approximately, the primary path P(z). Room impulse responses have delay and non-minimum-phase components, so such inversion is generally non-causal and ill-conditioned. The dereverberation and declipping experiments should either use c(n) = P(z)*s(n) as the target for all tasks, or the paper must justify how the active system realizes the required inverse filtering.","section":"§4.2"},{"comment":"The headline metric is directly optimized. The metric-based adversarial loss trains a discriminator to predict PESQ, and the final model is selected on validation PESQ. This creates a bias in favor of ASE-TM relative to the ANC baselines, which do not use a PESQ-based training or selection protocol. The paper should report an independent metric that was not used in training or model selection, or apply the same PESQ-guided training and validation procedure to the baselines, so that the PESQ comparisons are fair.","section":"§4.2 (LMetric) and §5.3"},{"comment":"The physical realizability of the active scheme is assumed rather than demonstrated. The simulation models the enhanced signal as the pointwise sum of a P(z)-filtered source and an S(z)-filtered loudspeaker output at a single modification microphone. In a real room, the loudspeaker radiates a spatially extended field and the microphone observes only one point of that field; the paper provides no experimental validation with measured RIRs or a real loudspeaker/microphone setup. The active-enhancement claim would be considerably strengthened by a real-world measurement or by a detailed discussion of the spatial and practical limitations of the point-to-point simulation.","section":"§5.2"}],"minor_comments":[{"comment":"The title contains typos: \"Decliping\" should be \"Declipping\" and \"Deveraberation\" should be \"Dereverberation.\" The abstract's \"outperforms existing baselines\" overstates the experimental scope; recommend \"outperforms the evaluated ANC baselines.\"","section":"Title/Abstract"},{"comment":"Equation citations are incorrect: the text says the anti-signal a(n) is \"as defined in Eq. 3\" and that eh(n) is \"as defined in Eq. 3,\" but Eq. (3) defines e(n)=d(n)−a(n), while a(n) is defined in Eq. (2) and eh(n) in Eq. (4).","section":"§4.1"},{"comment":"There are typos: \"respectivly\" should be \"respectively,\" and \"η2 = ∞\" in §6.4 should be \"λ2 = ∞.\" In §6.3, \"non-nonlinearities\" should be \"strong nonlinearities.\"","section":"§5.2, §6.4"},{"comment":"The relation between the task distortions in Eqs. (5)–(7) and the acoustic path model d(n)=P(z)*x(n) is not specified. For example, is the model input the degraded speech after the primary path, or is the degradation inserted before P(z)? Defining the complete signal flow for each task in a single equation block would make the simulations reproducible.","section":"§3 and §4.2"},{"comment":"The architecture diagram (Figure 2) omits the acoustic path blocks (P(z), S(z), fLS{·}) that are central to the ASE formulation; adding them would make the signal flow clearer. Figure 4 would benefit from a statement about whether the plotted spectra are averages over the test set or representative examples, since the caption says \"over the entire test set\" while the text could be read either way.","section":"Figures 2–4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope for audio and speech processing, but I would ask the editor to weigh whether the novelty disclosure is sufficient. The related work lists several active speech control papers (refs. [37], [93], [94]) but does not discuss how ASE differs from them in substance; in particular, the relationship to the authors' own Deep Active Speech Cancellation work (ref. [37]) should be clarified. More importantly, the current experiments do not yet support the paradigm-level claim, and the causality issue is not a minor editing point. If the authors can add the missing SE comparisons, enforce a physically valid causal budget, and redefine the dereverberation/declipping targets consistently, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a residual speech-enhancement model in ANC clothing. The narrow experiments are real and the ablation is informative, but the 'new paradigm' and real-time claims outrun the evidence, and Section 6.4 contradicts itself. The reader's conditional verdict is about right, and the stress-test note lands.\n\nWhat's actually there: the authors build a SEmamba-style Transformer-Mamba model (Mamba2 blocks plus one attention layer), train it with an MP-SENet-style multi-loss objective with L1 additions, and simulate primary/secondary acoustic paths with image-method RIRs and a saturating loudspeaker. They evaluate denoising, dereverberation, and declipping. Tables 1-3 show ASE-TM beating THF-FxLMS, DeepANC, and ARN on PESQ and NMSE. The ablation shows the loss changes, Mamba2, and attention contribute. That is legitimate experimental work and gives a referee something to check.\n\nSoft spots, in proportion:\n- The central framing is definitional. Eq. (4) defines eh = d + a and the model is trained to minimize distance to c. That is residual learning. Calling the correction an 'anti-signal' rendered by a loudspeaker is a physical story about deployment, not a new algorithmic task, and the physical story is only simulated.\n- Section 6.4 gives a causality budget of about 4.3 ms, then says the model predicts 500 future frames (31.25 ms) and calls that real-time. With the reference microphone only 2 m ahead, 31 ms of lookahead is not acoustically available. The main tables do not say whether the model is causal, so the reported gains may be offline and non-causal. This is the load-bearing weakness for the active claim.\n- The metric-adversarial loss targets PESQ and the validation checkpoint is selected on PESQ. That is a known and disclosed optimization of the headline metric; it does not invalidate the tables but it lowers their evidentiary value.\n- The abstract's 'outperforms existing baselines' is only supported against adapted ANC baselines. No standard SE model, including the SEmamba backbone, is compared. The authors list this as a limitation, but the abstract still overstates.\n- No code or data release, so the simulated-room pipeline is not independently checkable.\n\nWho this is for: readers working on deep ANC or on physical re-framings of speech enhancement. With causal experiments, a SEmamba baseline, and toned-down claims, this could be a useful paper. As is, I would not cite it, but I would send it to a serious referee: the simulation setup is non-trivial, the experiments are run carefully enough to be checked, and the fixes are concrete. Expect major revision.","headline":"Residual speech enhancement in ANC clothing: the narrow experiments are real and rerunnable, but the 'new paradigm' and real-time claims outrun the evidence, and Section 6.4 contradicts its own causality budget.","tokens_in":17827,"tokens_out":4630,"would_cite":false,"duration_ms":39712,"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 Active Speech Enhancement, in which a loudspeaker adds a learned correction signal to the degraded speech at the listener position, can outperform ANC-style cancellation filters across denoising, dereverberation, and…","keywords":["active speech enhancement","active noise cancellation","speech denoising","dereverberation","declipping","Transformer-Mamba","PESQ","room impulse response simulation"],"falsifier":"A hardware experiment in a real 3 x 4 x 2 meter room with the same microphone and loudspeaker geometry, using a measured loudspeaker transfer function instead of the Scaled Error Function, would settle the central claim: if the real-path ASE-TM does not beat the ANC baselines by the reported PESQ and NMSE margins, the result is a simulation artifact rather than a working active enhancement paradigm.","tokens_in":16805,"feed_emoji":"🎙️","tokens_out":5339,"duration_ms":30838,"temperature":0.7,"pith_summary":"The paper proposes replacing the goal of active noise cancellation with Active Speech Enhancement (ASE): instead of generating a signal that cancels an interfering noise, the system learns to add a shaped acoustic correction so that the sum at the error microphone is closer to clean speech. The authors introduce ASE-TM, a Transformer-Mamba network trained end-to-end with a weighted loss that combines time-domain, spectral, complex-spectrum, phase, metric-adversarial, and consistency terms. In simulated rooms with 512-tap image-method impulse responses and a Scaled Error Function loudspeaker nonlinearity, ASE-TM reports PESQ 2.98 for denoising, 2.43 for dereverberation, and 3.09 for declipping, clearly above the three adapted ANC baselines (THF-FxLMS, DeepANC, ARN). If correct, this would mean an active acoustics system can do more than cancel noise: it can restore a desired speech signal at a physical listening position. The authors note that the baselines were adapted from ANC, that the model is trained separately per task, and that the results rely on the simulated acoustic environment.","feed_headline":"Adding shaped sound outdoes active noise cancellation for speech","feed_subtitle":"Simulated-room tests show PESQ gains on denoising, dereverberation, and declipping.","key_machinery":"The load-bearing identity is the ASE superposition e_hat(n) = d(n) + a(n), where d(n) is the degraded signal arriving through the primary path P(z) and a(n) = S(z) * f_LS{y(n)} is the loudspeaker contribution through the secondary path S(z). The network, a Transformer-Mamba with Mamba2 blocks, an inserted multi-head attention block, and dense encoders and dual magnitude/phase decoders, is trained to output the complex spectrum of the loudspeaker signal y(n) under a combined loss L_G = gamma1 L_Time + gamma2 L_Mag + gamma3 L_Com + gamma4 L_Metric + gamma5 L_Pha + gamma6 L_Consist, with L_Time and L_Mag each combining L1 and L2 terms. This machinery converts speech enhancement from a filtering problem into a control problem: the optimizable object is an acoustic correction signal that physically recombines with the degraded signal, not a direct estimate of clean speech played to the listener.","core_discovery":"The central claim is that an active system with a reference microphone, a learned control filter, a loudspeaker, and a modification microphone can be trained so that the superposition e_hat(n) = d(n) + a(n) of the degraded primary signal and the loudspeaker anti-signal reconstructs a clean target speech signal c(n), rather than merely nulling the disturbance as in ANC. For additive noise the target is the clean speech after the primary path; for dereverberation and declipping the target is the original anechoic, unclipped utterance. The proposed ASE-TM predicts the complex spectrum of the loudspeaker input y(n), from which the anti-signal a(n) = S(z) * f_LS{y(n)} is formed and added to d(n). Across the three tasks, ASE-TM reaches PESQ 2.98, 2.43, and 3.09 respectively, with the ANC baselines staying below 2.45 in most conditions, and the paper presents this as evidence that active targeted modulation improves intelligibility and quality beyond cancellation.","pith_inferences":["Inference: If real rooms deviate from the 512-tap image-method RIRs and the Scaled Error Function model, the reported margins may shrink; a hardware prototype in a physical room of the simulated dimensions is the natural test of whether the paradigm transfers.","Inference: Defining the dereverberation and declipping targets as the original anechoic, unclipped speech implies the loudspeaker must compensate for the room transfer function itself, which is a more demanding control problem than canceling a disturbance and may explain why those tasks show the largest gains over ANC baselines.","Inference: Because NMSE and PESQ are computed against an unmeasurable clean target c(n) in real deployment, practical ASE systems would need a model of the room and the source or a separate reference signal, which could be a barrier the simulation does not expose.","Inference: A natural extension is a multi-condition ASE model that jointly handles noise, reverberation, and clipping, and, further out, personalizing the target c(n) to hearing-impaired listeners' audibility needs rather than exact clean speech."],"forward_implications":["If the ASE claim holds, active acoustic systems can be trained for joint interference suppression and signal enrichment, extending the scope of active control from noise removal to general speech restoration.","On the VoiceBank-DEMAND benchmark in the simulated room, ASE-TM exceeds the best ANC baseline by roughly 0.5 PESQ and improves NMSE to -21.76 dB, showing that the added correction does more than cancel the disturbance.","The same model and loss, with task-specific targets, transfers to dereverberation and declipping, where the ANC baselines often degrade the input rather than restore it, which the paper attributes to the difficulty of these tasks for cancellation-only methods.","With a future-frame prediction strategy, the model satisfies the causality condition T_ASE-TM < T_p - T_s (about 0.0043 seconds in the simulated geometry) and retains most of its quality, suggesting real-time operation is feasible.","A single model handling all three tasks simultaneously, rather than one trained per task, is left as future work."],"supporting_citations":[{"why":"Provides the SEmamba architecture and multi-level loss framework that ASE-TM extends, plus the evaluation protocol for PESQ, STOI, and composite measures.","marker":"[56]"},{"why":"Defines DeepANC and supplies the simulation setup (rectangular room, image-method RIRs, microphone and loudspeaker positions) that the paper adapts for ASE.","marker":"[31]"},{"why":"Defines the ARN attentive-recurrent-network baseline and the low-latency active control setting the paper compares against.","marker":"[54]"},{"why":"Defines the THF-FxLMS nonlinear adaptive baseline used as a classical ANC reference in all three tasks.","marker":"[71]"},{"why":"Supplies the VoiceFixer distortion definitions for additive noise, reverberation, and clipping, and the RIR training/test splits used for the dereverberation task.","marker":"[55]"},{"why":"Originates the multi-level magnitude/phase loss framework from which the ASE-TM optimization objective is built.","marker":"[63]"},{"why":"Provides the VoiceBank-DEMAND dataset used for training and evaluating the active denoising task.","marker":"[64]"},{"why":"The image method used to generate the room impulse responses for the primary and secondary acoustic paths in the simulation.","marker":"[68]"}],"fun_headline_variants":["Active Speech Enhancement outshines ANC on denoising, dereverb, declip","Shaped loudspeaker anti-signal reconstructs clean speech beyond ANC","Transformer-Mamba ASE system boosts PESQ on denoising, dereverb, declip","Active speech shaping reconstructs clean signal in noisy rooms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated room, with 512-tap image-method impulse responses for both acoustic paths and a Scaled Error Function loudspeaker saturation, faithfully represents a real listen-through environment and that e_hat(n) = d(n) + a(n) at the modification microphone is physically realizable.","fun_headline_variants_meta":{"raw":{"variants":["Active Speech Enhancement outshines ANC on denoising, dereverb, declip","Shaped loudspeaker anti-signal reconstructs clean speech beyond ANC","Transformer-Mamba ASE system boosts PESQ on denoising, dereverb, declip","Active speech shaping reconstructs clean signal in noisy rooms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000476,"raw_usage":{"total_tokens":2323,"prompt_tokens":872,"completion_tokens":1451,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1368}},"tokens_in":488,"tokens_out":1451,"duration_ms":9679,"temperature":1.0,"reasoning_tokens":1368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:52:31.743219+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A hardware experiment in a real 3 x 4 x 2 meter room with the same microphone and loudspeaker geometry, using a measured loudspeaker transfer function instead of the Scaled Error Function, would settle the central claim: if the real-path ASE-TM does not beat the ANC baselines by the reported PESQ and NMSE margins, the result is a simulation artifact rather than a working active enhancement paradigm.","supporting_citations":[{"cited_title":"VoiceFixer: A Unified Framework for High-Fidelity Speech Restoration","cited_arxiv_id":"2204.05841","evidence_quote":"Supplies the VoiceFixer distortion definitions for additive noise, reverberation, and clipping, and the RIR training/test splits used for the dereverberation task."}],"review_version":1}