{"id":"61492872-8a0e-4cef-85e1-115b3c9cfa7e","arxiv_id":"2412.18851","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An attention-gated short-time Wiener filter is fed as an extra feature into an ICCRN acoustic echo cancellation network, improving reported PESQ, SDR, and MOS ECHO scores.","lead":"The authors combine a classic Wiener filter formula with a learned attention gate as an extra input to a neural echo canceller, aiming to remove room echo without suppressing the speaker's own voice. On synthetic and blind test sets, the hybrid model reports better echo cancellation scores than several deep learning baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The attention-enhanced Wiener solution is not defined: no formula for V, no rule for splitting A into X^T X1 and X^T Y1, so Eqs. 9–14 do not yield a reproducible H_W and the reported gains cannot be traced to the described method.","rationale":"The reader's weakest_assumption identifies exactly the same gap: the construction of V and the decomposition of A into X^T X1 and X^T Y1 are unspecified, making the Wiener solution label non-functional. My reading of the paper confirms this: Section III-B provides no formula for V, and the transition from A to the two correlation matrices is asserted without derivation or dimensional analysis. This is the most load-bearing concern because it sits at the core of the proposed method; if the code does not fill this gap, the paper does not define a reproducible algorithm and the experimental results cannot be attributed to the described mechanism. The concern is addressable—the code is promised and the empirical comparisons appear plausible—so a conditional acceptance is appropriate rather than rejection. The reader's verdict of CONDITIONAL remains the correct one, and I therefore recommend no change. I found no additional internal inconsistency that would invalidate the empirical comparisons; the modification of MTFAA is noted but the primary improvement over the ICCRN baseline (the ablation comparison) is independent of that modification.","tokens_in":7520,"tokens_out":6854,"duration_ms":62223,"concrete_test":"Acquire the official code from https://github.com/ZhaoF-i/ASTWS-AEC. Trace the forward pass from X and Y to H_W: confirm that V is defined by an explicit formula in terms of X^T X and X^T Y, and that A is split deterministically into X^T X1 and X^T Y1 with shapes matching X^T X and X^T Y. Then re-run the synthetic test set at SER=0 and check whether the reported SDR of 23.46 dB is reproduced. If the formula is absent or the SDR differs substantially, the central claim is not supported by the paper as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central architectural step, the attention-enhanced Wiener solution, is underspecified to the point of non-implementability. In Section III-B, V is introduced as 'a combination of X^T X and X^T Y, as articulated in Equation 8,' but Equation 8 only states the Wiener normal equations; no explicit formula for V is given. Equation 11 merely gates V with a sigmoid. The attention output A is then asserted to be an enhanced representation of X^T X and X^T Y, 'denoted as X^T X1 and X^T Y1,' but the paper never states how A is decomposed into these two matrices, nor their shapes or dimensions. Consequently, the short-time Wiener step in Section III-C, which solves (X^T X1) H_W = X^T Y1, is not a well-defined function of X and Y. Moreover, the invertibility of X^T X1 is never addressed: no regularization, pseudo-inverse, or fallback is described, and H_W is not directly supervised since it is only used as an input to the ICCRN network. This leaves open the possibility that the empirically successful component is simply an extra attention-based feature, not the advertised Wiener solution, and prevents independent verification of the central claim that ASTWS integrates traditional filter theory.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents ASTWS, an acoustic echo cancellation system that combines an attention module with a short-time Wiener solution and an ICCRN-based regression network. The attention module is intended to emphasize single-talk frames and suppress double-talk interference before computing correlation-based filter coefficients; the resulting near-end estimate is fed alongside the far-end and microphone spectrograms into an ICCRN network. Experiments on a self-generated synthetic dataset and the ICASSP 2023 AEC Challenge blind test set report consistent gains over ICRN, MTFAA, and ICCRN, with an ablation (STWS vs. ASTWS) isolating the attention contribution. A code repository link is provided.","tokens_in":7828,"tokens_out":6222,"duration_ms":52604,"significance":"The paper addresses a relevant problem and, if the method is fully specified, the empirical results are potentially useful: ASTWS reports substantially higher double-talk SDR (e.g., 23.46 dB at SER=0 vs. 15.37 dB for MTFAA) with lower complexity than MTFAA, and also improves over its ICCRN backbone with modest overhead. The inclusion of a blind test set, an ablation study, and a code link are strengths. However, the central contribution—the attention-enhanced short-time Wiener solution—is described only at the level of a diagram and a few equations; the key construction V and the splitting of the attention output into correlation matrices are not defined. Until these are specified, the reported gains cannot be attributed to the advertised Wiener solution, and independent replication is impossible. The work is therefore not yet in a publishable form.","major_comments":[{"comment":"V is never defined. The text states only that V is 'a combination of X^T X and X^T Y, as articulated in Equation 8,' but no explicit formula is given, and the shapes of V, Q, and K after the gating transformations are not specified. Since V1 and hence A depend on this undefined quantity, Equations (9)-(13) do not define a reproducible attention output. Please give the exact construction of V (including dimensions and any normalization) and state the shapes of all intermediate tensors.","section":"Section III-B, Eqs. (9)-(13)"},{"comment":"The paper asserts that the attention output A is an enhanced representation of X^T X and X^T Y and 'denotes' these enhanced versions as X^T X1 and X^T Y1, but no rule is given for decomposing A into these two matrices. Without such a rule, solving (X^T X1) H_W = X^T Y1 is not a well-defined function of X and Y. This is the load-bearing step of the proposed method: it is the only connection between the attention mechanism and the Wiener solution, and it determines whether the label 'Wiener solution' is justified. Please provide the decomposition algorithm explicitly, or redefine the computation so that H_W follows from well-defined correlation matrices.","section":"Section III-B to III-C, Eqs. (13)-(14)"},{"comment":"The invertibility and numerical solution of the normal equations are not addressed. The matrix X^T X1 (or its regularized version) is not guaranteed to be invertible, yet no pseudo-inverse, diagonal loading, or fallback is described. In addition, H_W is not directly supervised—the loss in Equations (15)-(18) is applied only to the final network output—so it is unclear whether the reported improvements come from solving the Wiener equations or simply from feeding an extra learned feature to ICCRN. Please specify the linear solver and regularization, and consider an ablation that replaces H_W with a randomly initialized or unregularized correlation feature to demonstrate the specific contribution of the Wiener solution.","section":"Section III-C and Section III-D"}],"minor_comments":[{"comment":"Table I contains an unexplained row beginning 'mix 1.33 -10 1.93 0 2.56 10 ...' that appears to be a formatting artifact of the header; it should be removed or labeled.","section":"Table I"},{"comment":"Section V states that ASTWS achieves 'a greater performance improvement at a reduced computational expense relative to the ICCRN,' but Table I lists ASTWS at 0.963 GMacs / 0.148M parameters versus ICCRN at 0.844 GMacs / 0.12M parameters; the sentence should be corrected.","section":"Section V"},{"comment":"MTFAA is modified by removing its band decomposition and band merging modules; please clarify whether this is the standard configuration and discuss the effect of this modification on the reported baseline performance.","section":"Section IV-B"},{"comment":"The notation K is used both for the number of filter blocks in Eq. (2) and for the attention Key input in Eqs. (9)-(13); rename one of them to avoid ambiguity.","section":"Equations (2) and (9)-(13)"},{"comment":"Equation (15) defines the S-SISNR loss through beta and cos(beta) but does not state how beta is computed from the predicted and target signals; a reference or explicit formula is needed.","section":"Equation (15)"},{"comment":"The paper alternates between 'short-time Wiener solution' and 'short-term Wiener solution'; please use a single term consistently.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript links to a code repository, but the review was based on the text alone. Given that the main shortcoming is underspecification, the editor may wish to verify that the released code exactly implements Eqs. (9)-(14), including the construction of V and the splitting of A into X^T X1 and X^T Y1. If the code resolves the ambiguity, the authors should be asked to port those details into the paper; if not, the central claim is not verifiable. The self-citation of ICCRN by the same group is not itself a problem; the external comparisons to ICRN and MTFAA are appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate incremental extension of ICCRN — feed it a softmax-attention-gated Wiener-solution feature — and the reported gains over several baselines are plausible. The Wiener derivation itself is textbook and the empirical setup is clean. What does not hold up is the description of the attention-enhanced Wiener step: V is introduced only as 'a combination of X^T X and X^T Y' with no explicit formula; the split of the attention output A into X^T X1 and X^T Y1 is asserted without shapes or a rule; and the paper never says how X^T X1 is inverted or regularized. Equations 9–14 therefore do not define H_W as a reproducible function of X and Y. The stress-test note is right: this is not a nitpick, it is the load-bearing part of the claimed integration of classical filter theory. The ablation suggests the attention mechanism helps, but it could be helping as a generic feature extractor rather than as the advertised Wiener solution.\n\nWhat the paper does well: the comparisons are fair — same loss, same training strategy, same dataset — and the results are reported with MACs and parameter counts. ASTWS beats ICRN, MTFAA, and ICCRN on most metrics, with a particularly large SDR gain in double-talk. The MOS ECHO improvement on the ICASSP 2023 blind test set over MTFAA is substantial. The authors also promise code.\n\nSoft spots beyond the Wiener specification: no error bars, no multi-seed results, so the strong wording in the conclusions overreaches. This is common in the AEC literature but still worth noting.\n\nRecommendation: send it to peer review with the requirement that the authors supply the missing formulas, the shapes, the inversion details, and working code. If the method is effectively an attention-based feature, they should say so. This is fixable and the empirical results deserve checking. I would not cite it until the description is concrete and the code is verified.","headline":"A plausible incremental AEC model whose central attention-Wiener block is currently underspecified to the point of non-reproducibility; the empirical gains warrant refereed scrutiny if the authors fill in the missing math.","tokens_in":8298,"tokens_out":2685,"would_cite":false,"duration_ms":22775,"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":"An attention-enhanced short-time Wiener solution outperforms established acoustic echo cancellation models on synthetic and blind test sets.","keywords":["acoustic echo cancellation","short-time Wiener solution","attention mechanism","double-talk","ICCRN","speech enhancement","time-frequency domain","deep learning"],"falsifier":"Check whether the matrix A produced by the attention module can be split into the claimed enhanced auto-correlation and cross-correlation matrices and whether solving the resulting linear system reproduces the reported echo-reduced signal; if the split is arbitrary or the solve is inconsistent on held-out STFT frames, the Wiener explanation is not load-bearing.","tokens_in":7304,"feed_emoji":"🎧","tokens_out":10126,"duration_ms":72974,"temperature":0.7,"pith_summary":"This paper claims that a deep acoustic echo cancellation model can be improved by feeding it a classical Wiener filter estimate that has been cleaned by an attention mechanism. The proposed method, ASTWS, starts from the ICCRN network and adds an attention-refined short-time Wiener solution, whose echo-reduced output is passed to the network together with the far-end and microphone spectrograms. On a synthetic test set the paper reports higher perceptual speech quality (PESQ) and signal-to-distortion ratio (SDR) than the ICRN, MTFAA, and ICCRN baselines, especially in double-talk, and on the 2023 AEC challenge blind test set it reports the highest MOS ECHO score. The intended implication is that classical filter-theoretic knowledge and attention-based double-talk gating can be combined with a modern neural network at low computational cost.","feed_headline":"Attention-enhanced Wiener filter tops echo cancellation baselines","feed_subtitle":"Adding a short-time Wiener solution as a network input lifts PESQ, SDR, and MOS ECHO in double-talk tests.","key_machinery":"The central object is the attention-enhanced short-time Wiener solution: an attention module produces a matrix $A$ from gated versions of the far-end STFT signal, the microphone STFT signal, and a value matrix built from the correlations $X^TX$ and $X^TY$; the paper treats $A$ as supplying enhanced correlation matrices $X^TX_1$ and $X^TY_1$, solves $X^TX_1 H_W = X^TY_1$ for a filter $H_W$, and forms an echo-reduced signal $S_W[t,f] = D[t,f] - \\sum_{k=0}^{m-1} H_W[k,f]X[t-k,f]$. This $S_W$ is concatenated with the far-end and microphone STFT inputs to the ICCRN-based AEC module. The Wiener component is the mechanism that injects classical filter-theoretic structure; the attention component is what keeps that structure from being corrupted by double-talk.","core_discovery":"The central claim, stated on the paper's own terms, is that the attention-enhanced short-time Wiener solution (ASTWS) outperforms established baseline models on both a synthetic test set and the 2023 AEC challenge blind test set. The paper derives a short-time Wiener solution in the time-frequency domain, uses an attention mechanism to reduce the influence of double-talk on the auto- and cross-correlation terms, and feeds the resulting echo estimate into the ICCRN backbone. The reported results show consistent gains in PESQ and SDR across signal-to-echo ratios, with the largest improvement in double-talk SDR, and a higher MOS ECHO on the blind test set than the larger MTFAA model. The authors conclude that the approach proves both effectiveness and generalization for acoustic echo cancellation.","pith_inferences":["The paper leaves implicit how the value matrix $V$ is formed from $X^TX$ and $X^TY$ and how the attention output $A$ is split back into $X^TX_1$ and $X^TY_1$; making this split explicit and differentiable would let other architectures reuse the same input recipe.","A natural test of the mechanism is to freeze the Wiener branch and replace the attention weights with a random fixed gate; if performance is unchanged, the gain is coming from the backbone rather than from the attention-enhanced Wiener estimate.","The same attention-gated filter-theoretic input strategy may transfer to other classical estimators such as recursive least squares or Kalman filtering, where double-talk robustness is also the main failure mode.","Exposing the filter coefficients $H_W$ themselves to the network, rather than only the echo-reduced signal $S_W$, might let the network correct filter misalignments and improve double-talk performance further."],"forward_implications":["If the reported results hold, ASTWS becomes a stronger and cheaper alternative to ICRN and MTFAA, improving double-talk SDR by several decibels at a fraction of the parameter count.","The success of the short-time Wiener input suggests that classical filter-theoretic priors can be injected into neural AEC models as input features rather than as replacement filters, which is a general recipe for combining old and new methods.","The large SDR jump from the no-attention ablation (STWS) to the full method (ASTWS) indicates that the attention module, not just the Wiener input, is a direct contributor to double-talk performance.","Because the Wiener solve is causal and fixed-length with $m=20$ blocks, the input strategy is compatible with streaming, low-latency deployment.","The improvement on the 2023 blind test set indicates the gain is not tied to the specific synthetic room settings used in training."],"supporting_citations":[{"why":"Defines the ICRN baseline and supplies the nonlinearity configuration used to generate the synthetic training data.","marker":"[8]"},{"why":"Defines the MTFAA baseline that ASTWS compares against and exceeds in MOS ECHO on the blind test set.","marker":"[10]"},{"why":"Provides the ICCRN backbone network whose single-layer codec the proposed AEC module extends.","marker":"[16]"},{"why":"Provides the synthetic dataset and the blind test set used for training and evaluation.","marker":"[11]"},{"why":"Defines the S-SISNR loss component included in the total training objective.","marker":"[18]"},{"why":"Generates the room impulse responses used to simulate training rooms and echo paths.","marker":"[20]"}],"fun_headline_variants":["Short-time Wiener attention boosts echo cancellation","Attention Wiener filter beats AEC baselines","Wiener-guided attention improves AEC double-talk","ASTWS: attention Wiener outperforms AEC models","Wiener attention ups echo cancel gains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the attention output can be separated into an enhanced auto-correlation matrix and an enhanced cross-correlation matrix, and that solving the resulting linear system yields a filter whose echo-reduced signal genuinely helps the neural network.","fun_headline_variants_meta":{"raw":{"variants":["Short-time Wiener attention boosts echo cancellation","Attention Wiener filter beats AEC baselines","Wiener-guided attention improves AEC double-talk","ASTWS: attention Wiener outperforms AEC models","Wiener attention ups echo cancel gains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1179,"prompt_tokens":845,"completion_tokens":334,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":267}},"tokens_in":461,"tokens_out":334,"duration_ms":3217,"temperature":1.0,"reasoning_tokens":267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:24:40.582182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Check whether the matrix A produced by the attention module can be split into the claimed enhanced auto-correlation and cross-correlation matrices and whether solving the resulting linear system reproduces the reported echo-reduced signal; if the split is arbitrary or the solve is inconsistent on held-out STFT frames, the Wiener explanation is not load-bearing.","supporting_citations":[{"cited_title":"A complex spectral mapping with inplace convolution recurrent neural networks for acoustic echo cancellation,","cited_arxiv_id":null,"evidence_quote":"Defines the ICRN baseline and supplies the nonlinearity configuration used to generate the synthetic training data."},{"cited_title":"Multi-scale temporal frequency convolutional network with axial atten- tion for speech enhancement,","cited_arxiv_id":null,"evidence_quote":"Defines the MTFAA baseline that ASTWS compares against and exceeds in MOS ECHO on the blind test set."},{"cited_title":"Iccrn: Inplace cepstral convolutional recurrent neural network for monaural speech enhancement,","cited_arxiv_id":null,"evidence_quote":"Provides the ICCRN backbone network whose single-layer codec the proposed AEC module extends."},{"cited_title":"Funnel deep complex u-net for phase-aware speech enhancement,","cited_arxiv_id":null,"evidence_quote":"Defines the S-SISNR loss component included in the total training objective."},{"cited_title":"Image method for efficiently simulating small-room acoustics,","cited_arxiv_id":null,"evidence_quote":"Generates the room impulse responses used to simulate training rooms and echo paths."}],"review_version":1}