{"id":"f9dc0887-95c2-4bff-bdca-2574597e352c","arxiv_id":"2501.16367","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified benchmark shows DNN-supported Kalman filters for acoustic echo cancellation converge faster and cancel more echo than the classical FDKF, with per-bin methods best preserving near-end speech.","lead":"This paper compares four deep-learning-augmented Kalman filter variants for acoustic echo cancellation head-to-head under one training setup. It finds that the neural variants converge faster and cancel echo better than the classical frequency-domain Kalman filter, with per-bin designs preserving near-end speech best.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No error bars or significance tests are reported; the paper explicitly states large standard deviations, so the central claim that neural Kalman filters outperform FDKF is not statistically established.","rationale":"The paper is a well-structured comparative study with transparent code and a consistent evaluation framework, which is a strength. However, the core assertion of superiority is comparative, and the paper's own decision not to report standard deviations undermines the reliability of every ranking in Fig. 7 and the convergence curves. The reader's concern about re-implementation fidelity is relevant to external validity, but the statistical gap is more fundamental: even if all implementations were faithful, the observed means might be within noise. I therefore see the missing significance testing as the most load-bearing concern. A paired test over the 60 files would settle it without new data. If the ERLE margins are as large as they appear, significance may hold; the near-end speech metrics (PESQ/STOI) with smaller margins are more at risk, and those directly support the claim of equal or better NE preservation. Thus the condition on the paper should be to add statistical analysis and to temper the categorical conclusion based on the WGN results.","tokens_in":21198,"tokens_out":12293,"duration_ms":125496,"concrete_test":"For each of the 60 test files in Dtest, DWGN, and DNL, compute per-file ERLE, PESQ, STOI, LPS, and AECMOS scores for FDKF and each neural method. Run paired significance tests (e.g., Wilcoxon signed-rank) for every method-vs-FDKF pair and metric, with a multiple-comparison correction (e.g., Benjamini-Hochberg). Report median/quartiles and effect sizes. If a neural method fails to significantly beat FDKF on the primary echo metric in a condition, qualify the conclusion accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim that 'neural Kalman filters reveal better echo performance than FDKF along with better (re)convergence' rests entirely on mean metric comparisons over 60 test files per condition, yet the paper explicitly states: 'the broad range of SER values naturally leads to large standard deviations in most metrics. Accordingly, we do not report standard deviations.' Without error bars, confidence intervals, or significance tests, the observed differences (e.g., DLAC-Kalman's small PESQ/STOI lead over FDKF, or the 10 dB ERLE gaps in Figs. 4–5) cannot be distinguished from noise. The problem is compounded by the conclusion's categorical wording: under WGN excitation (Fig. 6), DeepAdaptive and NeuralKalman show ERLE limited to about 10 dB while FDKF reaches a 'decent final accuracy', so the claim is not even uniformly supported by the paper's own results. Any comparative ranking and the derived design recommendations (per-bin vs. fully connected, etc.) require statistical backing before they can be considered established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper revisits the frequency-domain adaptive Kalman filter (FDKF) for acoustic echo cancellation and presents a unified framework for four neural Kalman filter variants (DLAC-Kalman, NKF, NeuralKalman, DeepAdaptive) that replace or augment components such as Kalman gain, state transition, distortion modeling, or filter-state update with DNNs. It then reports a comparative evaluation against FDKF and an end-to-end neural baseline (CGGN16) under linear and nonlinear loudspeaker conditions, including double-talk, RIR switches, and WGN excitation. The central claims are that neural Kalman filters achieve faster (re)convergence, better echo cancellation, and in some cases better near-end speech preservation than FDKF, and that the comparison provides design guidance (e.g., per-bin vs. fully connected processing).","tokens_in":21344,"tokens_out":2666,"duration_ms":37820,"significance":"If the central claims are established, this is a valuable contribution: it provides a common mathematical notation and experimental framework for a recently emerging class of hybrid DNN/Kalman AEC methods, includes a careful attempt to equalize algorithmic delay and effective reference length, and releases code and evaluation scripts. The inclusion of an Oracle-FDKF upper bound, a fully data-driven CGGN16 baseline, complexity and parameter counts, and a language-independent LPS metric are commendable and strengthen the paper's utility as a benchmark. However, the comparative ranking that carries the conclusions is not statistically supported because the authors explicitly omit standard deviations, and the re-implementations of prior methods involve nontrivial modifications, so the significance of the specific rank order remains to be established.","major_comments":[{"comment":"The paper states on p. 14 that 'the broad range of SER values naturally leads to large standard deviations in most metrics. Accordingly, we do not report standard deviations.' This is load-bearing for the central claim because the ranking of methods is based on mean metric differences over 60 files per condition, and Fig. 7 presents bar charts with no error bars or significance tests. For example, the PESQ/STOI/LPS differences between DLAC-Kalman and FDKF in Fig. 7 appear small, and without confidence intervals the observed ordering cannot be distinguished from sampling noise. I request the authors to either report standard deviations or confidence intervals, or perform pairwise significance tests (e.g., matched-pairs bootstrap on the 60 files), and to adjust the conclusions to the strength of the statistical evidence.","section":"§IV-C, Fig. 7"},{"comment":"The manuscript states on p. 9 that the authors 'replicate the authors' methods as closely as possible,' but then lists departures: NeuralKalman is modified to use a 3x3 causal kernel, DLAC-Kalman is implemented in a narrowband variant, all models are retrained from scratch in a common PyTorch framework with the same loss, and DFT sizes and filter lengths are changed (K=512, 1408, 896) to equalize delay and effective reference length. These changes mean the reported ranks and design recommendations apply to the authors' re-implementations rather than necessarily to the original published methods. The paper should either validate the re-implementations against the original code/results or explicitly state in the conclusions that the ranking is specific to the re-implemented versions used here.","section":"§III-A–D and §IV-A"},{"comment":"The conclusion that 'neural Kalman filters reveal better echo performance than FDKF along with better (re)convergence' is contradicted by the paper's own WGN results: in Fig. 6, DeepAdaptive and NeuralKalman are limited to roughly 10 dB ERLE while FDKF reaches a much higher final ERLE, and NKF/DLAC-Kalman also appear close to FDKF. The claim is therefore not uniformly supported across test sets. The conclusions should be qualified to the speech-excitation conditions, or the paper should provide a mechanism (e.g., an interaction analysis of method × excitation) that reconciles this condition with the overall claim.","section":"§IV-D, Fig. 6; §V"}],"minor_comments":[{"comment":"The 'informal subjective listening' section reports impressions from 106 files but provides no details on the number of listeners, the rating scale, or any statistical treatment; it is anecdotal and should be labeled as such rather than being used to support the comparative ranking.","section":"§IV-F"},{"comment":"The name 'DeepAdaptive' is occasionally written as 'Deep Adaptive' (e.g., p. 16, Fig. 5 caption); please standardize the spelling.","section":"Throughout"},{"comment":"The observation noise variance smoothing factor β=0.5 is introduced without comment; a sentence or citation explaining this choice would improve reproducibility.","section":"Eq. (13)"},{"comment":"The definition of the effective reference input length M = K + (L−1)·R is helpful, but the notation R is used both for frame shift and for RIR length in the introduction (Eq. (1) uses N for RIR length); please ensure the symbols do not clash in the reader's mind.","section":"§IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is a useful systematic comparison and the authors have made commendable efforts to equalize delay and context across methods. The main risk is that the headline conclusions are stated categorically despite the explicit absence of variance reporting. This is fixable within the manuscript's scope by adding error bars/significance tests or by softening the claims, so I do not think rejection is warranted. I would also encourage the editor to ask the authors to clarify the status of re-implementations relative to the original methods, since that determines whether the paper is a benchmark of published algorithms or of the authors' re-implementations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this paper as a benchmark and synopsis, not as a new algorithm paper. The authors standardize four published neural FDKF extensions into one PyTorch framework with matched delay, effective reference length, training data, and loss, then compare them against classical FDKF on linear and nonlinear loudspeaker conditions. That common-framework comparison is genuinely new and useful. The unified notation in Table I and Fig. 2 makes architectural differences explicit, and the observation that NeuralKalman and DeepAdaptive behave like mask-based echo suppressors rather than subtractive cancellers is a real insight the original papers did not spell out. They also report complexity and parameter counts and discuss per-bin versus fully connected tradeoffs, which practitioners will actually use. The code and metric toolbox are public.\n\nThe soft spot is the one the stress-test note flags: there are no error bars or significance tests anywhere, and the authors say so explicitly: \"the broad range of SER values naturally leads to large standard deviations in most metrics. Accordingly, we do not report standard deviations.\" Since the central claim that neural Kalman filters outperform FDKF in echo reduction and reconvergence rests on mean differences over 60 files per condition, the categorical conclusion overreaches. The WGN results in Fig. 6 even show DeepAdaptive and NeuralKalman capped around 10 dB ERLE while FDKF reaches decent accuracy, so the claim is not uniformly supported by the paper's own data. That does not destroy the paper, but it means the ranking and design recommendations are directional, not definitive.\n\nA second concern is re-implementation fidelity. The authors say they replicated methods as closely as possible, but they modified NeuralKalman's kernel to a 3x3 causal kernel, selected a narrowband DLAC-Kalman variant, retrained everything from scratch with the same loss, and changed DFT sizes and taps to equalize delay. Those are reasonable choices for fairness, but they make it hard to know whether the observed ranking transfers to the actual published methods. To their credit, the authors disclose all of this.\n\nThe FDKF recap and the conceptual \"state as bridge\" discussion are fine but not deep; this is not a theory paper. The informal listening section is honest, and the DTO metric caveat is well handled.\n\nBottom line: send it to review, conditional on the authors adding variance information or significance testing and documenting re-implementation fidelity more thoroughly. It deserves a serious referee; the comparison itself is valuable enough. I would cite it as a benchmark and bring it to reading group as a good example of how to run a fair cross-method evaluation.","headline":"Useful controlled comparison of neural Kalman AEC variants, but the headline claim overreaches what the statistics support.","tokens_in":21893,"tokens_out":1755,"would_cite":true,"duration_ms":18389,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Neural Kalman filters beat classic FDKF on echo and reconvergence.","keywords":["acoustic echo cancellation","frequency-domain adaptive Kalman filter","neural Kalman filter","double-talk","deep learning","echo return loss enhancement","system identification"],"falsifier":"Run the four original released implementations (or author-provided code) with their original filter lengths, DFT sizes, losses, and training procedures on the paper's Dtest and DNL test sets and compare ERLE and (re)convergence against FDKF; if FDKF matches or beats the neural variants on echo reduction or double-talk near-end measures under those settings, the paper's ranking and design recommendations would not transfer to the published methods.","tokens_in":1770,"feed_emoji":"🎤","tokens_out":1736,"duration_ms":61471,"temperature":0.7,"pith_summary":"This paper claims that replacing parts of the frequency-domain adaptive Kalman filter (FDKF) with small neural networks yields faster convergence and reconvergence, stronger echo cancellation, and comparable or better near-end speech preservation during double-talk, both for linear and nonlinear loudspeakers. The authors unify four recent hybrid designs in a common state-space notation, retrain them from scratch in one framework on identical data, and evaluate them against the classical FDKF and a fully data-driven baseline. They report that the hybrid methods outperform FDKF in echo-return loss enhancement and reconvergence after room-impulse-response changes, while per-bin designs (DLAC-Kalman and NKF) preserve near-end speech best. If true, the practical takeaway is that a small network estimating the Kalman gain can improve the model-based FDKF without sacrificing its double-talk robustness, and that fully connected hybrids lose to end-to-end networks.","feed_headline":"Neural Kalman filters beat classic FDKF on echo and reconvergence","feed_subtitle":"One training setup: neural Kalman filters improve echo reduction and reconvergence over the FDKF baseline.","key_machinery":"The central object is the acoustic state-space model, a first-order Markov process $h(n+1)=a\\,h(n)+\\Delta h(n)$ for the room impulse response combined with the observation $y(n)=s(n)+n(n)+x^{\\mathsf T}(n)h(n)$, whose optimal recursive estimator is the frequency-domain adaptive Kalman filter. The FDKF updates a per-bin filter state $\\hat{H}_{\\ell,k}=A\\hat{H}_{\\ell-1,k}+A K_{\\ell,k} E_{\\ell,k}$ with a Kalman gain $K_{\\ell,k}$ that balances echo-path tracking against observation noise; its practical weakness is that the process and observation covariances must be supplied a priori. The machinery is to keep this update loop intact and replace individual blocks with DNNs -- the Kalman gain, the transition factor $A_\\ell$, the filter-state postprocessing, or the reference signal (distortion model) -- and the comparison is made meaningful by training all variants in one framework with the same log-MSE echo loss and equalized algorithmic delay.","core_discovery":"The central claim is that the classical FDKF's limitations -- model linearity and the need to hand-set noise covariances -- can be relieved by letting DNNs take over one of four roles: estimating the Kalman gain, estimating the state-transition factor, postprocessing the filter-state update, or modeling loudspeaker distortion as a learned reference signal. Under a fair common framework (same delay, same effective reference length, same data and loss), each of the four neural Kalman filters reaches higher ERLE and faster (re)convergence than the FDKF on linear and nonlinear echo paths, and the gain-estimating variants match or exceed the FDKF's double-talk near-end preservation as measured by PESQ, STOI, and LPS. The authors further observe that methods with a learned distortion model behave like mask-based echo suppressors rather than subtractive cancellers, which explains their aggressive echo reduction but weaker near-end quality; per-bin processing with shared weights gives smaller parameter counts and flexible filter lengths, while fully connected hybrids are outpaced by an end-to-end network in most metrics and resources.","pith_inferences":["The paper's equalization of algorithmic delay and effective reference length, though fair, may favor OLS-based FDKF and penalize OLA-based hybrids as originally tuned; if original authors' settings change results, the ranking is implementation-sensitive.","Because the trained distortion models can absorb the echo path (e.g., learning $\\hat{H}=1$), the Kalman loop may decouple from the physical echo path; a testable extension is to constrain the distortion model with an auxiliary loss to keep the filter identifiable.","The per-bin weight-sharing designs suggest a path to few-shot or variable-length AEC, where one compact network trained at a fixed DFT size operates at different filter lengths, an extension the paper notes but does not test.","The observation that the AECMOS 'other degradation' metric underrates near-end distortion at low SER suggests that objective metrics may mislead hybrid-AEC comparisons; an ASR-based or listening-test evaluation on the paper's test sets would strengthen the near-end preservation ranking."],"forward_implications":["If the central claim is right, a small DNN estimating the Kalman gain (DLAC-Kalman or NKF) can improve FDKF's echo reduction and convergence without harming near-end speech, making it a practical low-parameter upgrade.","The learned-distortion hybrids (NeuralKalman, DeepAdaptive) behave as mask-based echo suppressors, so they should be evaluated and deployed with that behavior in mind, not as pure system identifiers.","The FDKF remains the minimal-complexity choice when compute or parameters are the binding constraint, while fully connected neural Kalman filters are dominated by end-to-end networks in most metrics and resources.","Per-bin architectures decouple frame rate from update rate and can reuse one trained model at different filter lengths, which suits low-latency or variable-latency operation.","Residual echo from nonlinearities and long echo tails remains, so a postfilter is still beneficial after neural Kalman filters."],"supporting_citations":[{"why":"Defines the FDKF baseline and the acoustic state-space model that the hybrid methods extend.","marker":"[5]"},{"why":"Supplies the process noise covariance formula used in the FDKF and the synthetic RIR generation used for training.","marker":"[19]"},{"why":"Documents FDKF's near-end speech preservation in double-talk, the behavior the neural variants aim to match or exceed.","marker":"[23]"},{"why":"Introduces DLAC-Kalman, the narrowband DNN-based Kalman gain estimator evaluated here.","marker":"[32]"},{"why":"Introduces NKF, the per-bin neural Kalman filtering architecture evaluated here.","marker":"[33]"},{"why":"Introduces NeuralKalman, with DNN-based distortion, transition, and filter-state postprocessing, evaluated here.","marker":"[34]"},{"why":"Introduces DeepAdaptive, the hybrid with DNN-based gain and distortion modeling, evaluated here.","marker":"[35]"},{"why":"Provides the AECMOS metric used to assess echo reduction and near-end degradation.","marker":"[37]"}],"fun_headline_variants":["Neural Kalman filters outdo FDKF in echo cancellation and reconvergence","DNN-assisted Kalman filters improve echo reduction and reconvergence","Four neural Kalman filters surpass FDKF in acoustic echo tests","Neural Kalman filters: faster reconvergence, better echo cancellation"],"cache_read_input_tokens":24064,"weakest_assumption_plain":"The ranking assumes that the authors' re-implementations of DLAC-Kalman, NKF, NeuralKalman, and DeepAdaptive faithfully represent the original published methods, since every model was retrained from scratch with a common loss, modified filter lengths, DFT sizes, and (for NeuralKalman) a changed kernel, any of which could alter the outcome.","fun_headline_variants_meta":{"raw":{"variants":["Neural Kalman filters outdo FDKF in echo cancellation and reconvergence","DNN-assisted Kalman filters improve echo reduction and reconvergence","Four neural Kalman filters surpass FDKF in acoustic echo tests","Neural Kalman filters: faster reconvergence, better echo cancellation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000324,"raw_usage":{"total_tokens":1829,"prompt_tokens":970,"completion_tokens":859,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":781}},"tokens_in":586,"tokens_out":859,"duration_ms":8407,"temperature":1.0,"reasoning_tokens":781,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:31:12.549393+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the four original released implementations (or author-provided code) with their original filter lengths, DFT sizes, losses, and training procedures on the paper's Dtest and DNL test sets and compare ERLE and (re)convergence against FDKF; if FDKF matches or beats the neural variants on echo reduction or double-talk near-end measures under those settings, the paper's ranking and design recommendations would not transfer to the published methods.","supporting_citations":[{"cited_title":"Frequency-Domain Adaptive Kalman Filter for Acoustic Echo Control in Hands-Free Telephones,","cited_arxiv_id":null,"evidence_quote":"Defines the FDKF baseline and the acoustic state-space model that the hybrid methods extend."},{"cited_title":"An Automotive Wideband Stereo Acoustic Echo Canceler Using Frequency- Domain Adaptive Filtering,","cited_arxiv_id":null,"evidence_quote":"Supplies the process noise covariance formula used in the FDKF and the synthetic RIR generation used for training."},{"cited_title":"An Efficient Residual Echo SupressionforMulti-ChannelAcousticEchoCancellationBased on the Frequency-Domain Adaptive Kalman Filter,","cited_arxiv_id":null,"evidence_quote":"Documents FDKF's near-end speech preservation in double-talk, the behavior the neural variants aim to match or exceed."},{"cited_title":"End-to-End Deep Learning-Based Adaptation Control for Linear Acoustic Echo Cancellation,","cited_arxiv_id":null,"evidence_quote":"Introduces DLAC-Kalman, the narrowband DNN-based Kalman gain estimator evaluated here."},{"cited_title":"Low- Complexity Acoustic Echo Cancellation with Neural Kalman Filtering,","cited_arxiv_id":null,"evidence_quote":"Introduces NKF, the per-bin neural Kalman filtering architecture evaluated here."},{"cited_title":"DeepAdaptiveAEC:HybridofDeepLearning and Adaptive Acoustic Echo Cancellation,","cited_arxiv_id":null,"evidence_quote":"Introduces DeepAdaptive, the hybrid with DNN-based gain and distortion modeling, evaluated here."},{"cited_title":"AECMOS: A Speech Quality Assessment Metric for Echo Impairment,","cited_arxiv_id":null,"evidence_quote":"Provides the AECMOS metric used to assess echo reduction and near-end degradation."}],"review_version":1}