{"id":"409b9b1b-069a-4bfe-b99c-21dac38a14fd","arxiv_id":"2502.06098","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A delay-estimation classifier trained on adaptive filter bank energies improves time delay accuracy and speech enhancement in synthetic echo cancellation tests.","lead":"This paper proposes an echo cancellation system that estimates audio delay using a bank of adaptive filters whose weight energies are scored by a small neural network classifier. The system then suppresses residual echo with another neural network plus OMLSA and applies automatic gain control with smoothed gains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The TDE claims rely on test RIRs whose direct paths are all aligned to sample 0 (Sec. 3.1), so the only delay the classifier must learn is the artificial uniform shift; no evidence shows it handles real echo-path delays.","rationale":"The reader's weakest assumption precisely identifies the direct-path-alignment issue, and my stress-test confirms that this is the most load-bearing threat to the central claim. The paper's reported accuracies are internally plausible, but they are produced by a closed synthetic loop in which the only delay to estimate is an artificial global shift. A real AEC system must estimate an unknown delay embedded in the acoustic echo path, and the evaluation never tests that condition. Because this is a missing-evidence concern rather than a demonstrated failure, the existing CONDITIONAL verdict is appropriate: the paper needs a real-room or unaligned-RIR evaluation and public implementation before the central claims can be accepted. No verdict adjustment is needed if the reader's conditions are retained.","tokens_in":6581,"tokens_out":5647,"duration_ms":57719,"concrete_test":"Keep the trained classifier fixed, but generate a new TDE test set with the alignment step removed: leave each RIR at its original direct-path position, set the artificial delay to zero (or draw delays from real-room distributions), and recompute the Table 1 accuracies. If accuracy drops sharply or falls below the WebRTC baselines, the alignment assumption is load-bearing; if accuracy remains above 90%, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing uncertainty is whether Table 1's 91.67% (±25 ms) and 89.88% (±5 ms) accuracies describe real acoustic delay estimation or only a synthetic artifact. Section 3.1 states that 'the begin of the direct path was identified and set to position 0' for every RIR. In the TDE training-generation procedure (Fig. 2(a)), the full near-end mixture is then delayed by an artificially chosen τ. Thus the adaptive-filter energy patterns the classifier sees are generated by a global uniform shift of the microphone signal, not by the physical reference-to-microphone delay that AEC needs; the RIR itself contributes no delay. In a real system, the direct-path delay is unknown, varies with room geometry and hardware latency, and is precisely the quantity to be estimated. This protocol gives the classifier an unrealistically clean, single-cause delay. The same synthetic distribution is used for the AEC results in Table 2, so the mean PESQ gain of 0.633 also does not by itself establish real-world transfer. This is a missing-evidence concern rather than an internal contradiction: the measured numbers may be correct on the defined synthetic task, but they do not yet support the stronger claim that the scheme is robust for actual TDE/AEC. Secondary issues, such as the 'M×N=532' inconsistency in Sec. 2.1 and the absence of released code, reinforce the need for verification but are not the central risk.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hybrid AEC system with four components: (i) a time-delay estimator built from a bank of five multidelay block adaptive filters whose per-block weight energies are classified by a lightweight neural network; (ii) a residual-echo and noise suppression network based on RNNoise-style MFCC features with a double-talk detection output; (iii) an OMLSA postfilter that treats residual echo as a special noise; and (iv) an automatic gain control with sigmoid-based gain smoothing. Experiments on a synthetic dataset derived from the AEC Challenge 2021 and DNS-Challenge corpora report TDE accuracies of 91.67 percent within 25 ms and 89.88 percent within 5 ms, and a mean PESQ gain of 0.633 for the complete processing chain, outperforming the reported WebRTC and DTLN baselines.","tokens_in":7116,"tokens_out":4626,"duration_ms":43985,"significance":"The central idea of using the energy pattern of an adaptive-filter bank as a feature for delay classification is plausible and well aligned with current hybrid DSP/neural AEC research. The paper also provides useful implementation details, including filter-bank geometry, training data generation, model sizes, and comparisons with representative baselines. If the reported TDE accuracy and PESQ gains were shown to transfer to realistic acoustic echo paths, the contribution would be solid. However, the empirical evidence as presented is entirely in-distribution on a synthetic pipeline in which the direct-path delay of every room impulse response is removed; this is a load-bearing gap for the paper's main claim. The lack of statistical testing and the arithmetic inconsistency in the network dimensions further reduce confidence.","major_comments":[{"comment":"Section 3.1 states that for every room impulse response 'the begin of the direct path was identified and set to position 0'. The TDE training data are then created by delaying the resulting microphone mixture by an artificially chosen tau. This means the only delay present in the adaptive-filter features is the imposed uniform shift; the physical direct-path delay of the echo path, which is precisely the quantity a real time-delay estimator must handle, is absent. Table 1 and Table 2 are computed on this aligned synthetic pipeline, so the reported accuracy and PESQ gains do not support the conclusion that the method is robust for real TDE/AEC. The authors should add experiments on unaligned RIRs with varied direct-path delays, or on real recorded data, and report the corresponding accuracies and PESQ gains.","section":"3.1"},{"comment":"In Section 2.1, the stated parameters N=32, M=5 give M x N = 160, not 532 as printed. With L=8, the number of buffer positions and classifier categories is M x N - L = 152, not 155. Since the text relies on these numbers to define the classifier input dimension and output categories, the arithmetic error makes the architecture ambiguous. Please correct the values and state the exact input and output dimensions used in the implementation.","section":"2.1"},{"comment":"The evaluation reports point estimates without confidence intervals or significance tests. In Table 2, the PESQ-gain variance values (0.101 for +++OMLSA, 0.162 for ++NN) are large relative to the differences in mean gains between the proposed scheme (0.633) and the End2End baseline (0.545). Without the number of test clips, error bars, or paired statistical tests, the claimed superiority over the baselines is not established. Please provide these statistics or soften the comparative claims accordingly.","section":"3.2"}],"minor_comments":[{"comment":"The text says the TDE data are prepared 'as is illustrated in Fig. 3(a)', but the correct reference is Fig. 2(a); the same inconsistency affects the figure reference for the suppression network data preparation in Section 2.2.","section":"3.1"},{"comment":"The notation for the speech-present probability is inconsistent between Eqs. (2) and (3): it appears as p'(k,l) and p(k,l), and also as p(l,k). Please unify the notation and define each symbol at first use.","section":"2.2"},{"comment":"The sentence 'As is is shown in Fig. 1' contains a duplicated word, and Section 3.2 refers to 'OLMSA' instead of 'OMLSA'.","section":"2.1"},{"comment":"The text states that all components can be realized from open-sourced code, but no code, trained models, or configuration files are released. A statement about the planned release would strengthen reproducibility.","section":"5"}],"recommendation":"major_revision","confidential_remarks":"The central barrier is the mismatch between the synthetic evaluation and the real-world TDE problem: direct-path alignment artificially removes the very delay that the method is supposed to estimate. I do not think this is an internal inconsistency in the measured numbers, but it is a missing-evidence issue that is load-bearing for the main claim. The authors can address it within the manuscript's scope by testing on unaligned RIRs or real device data, so I recommend major revision rather than rejection. The arithmetic error in Section 2.1 should also be corrected before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the adaptive-filter-bank delay estimator is a genuinely new twist, and the system is assembled from sensible components, but the evaluation currently aligns the RIR direct path to zero, which removes the exact delay a real TDE would have to find—so the 91.67% accuracy and 0.633 PESQ gain are measured on an artificially clean version of the task.\n\nWhat's actually new: using the block-energy profile of a bank of overlapping adaptive filters as input features to a small classifier for delay estimation. That is not in the prior TDE literature I know. The idea is cheap, plausible, and the paper honestly credits WebRTC AEC3 and RNNoise for the architecture. The AGC smoothing with a sigmoid cross-fade is also a reasonable engineering touch. The paper is clearly written and follows a sensible train/eval split.\n\nSoft spots: first, the RIR alignment issue. Section 3.1 says the direct path of every RIR is identified and set to position 0. That means the only delay the classifier sees is the artificial τ added to the mixture. In a real AEC, the reference-to-microphone delay includes the loudspeaker-to-microphone propagation plus hardware latency, and it varies. No evidence here that the filter-bank features capture that unknown delay. Table 1 therefore describes a synthetic subproblem, not the problem AEC needs solved. Second, the internal numbers don't add up: N=32, M=5 gives M×N=160, not 532, and the category count M×N−L is also inconsistent. Third, no significance tests, no error bars, and the PESQ variance in Table 2 is large relative to the mean gains. Fourth, the suppression network is trained on the same generated distribution and uses the estimated delay to create the error signal, so the AEC gains are also in-distribution.\n\nNone of this means the method is wrong. The measured numbers are probably correct on the defined synthetic task. But the paper overclaims robustness, and the missing pieces—public code, a real RIR delay setup, or at least an evaluation with unaligned RIRs—are essential before the results mean much for practice.\n\nMy recommendation: send it to peer review, but clearly flag that the TDE evaluation needs to be redone with unaligned RIRs or real recordings, and the arithmetic errors need fixing. A competent referee can sort this out in one round. I wouldn't cite it in my own work until that verification exists.","headline":"Novel filter-bank TDE idea, but the synthetic evaluation aligns away the real delay, so the headline numbers don't yet transfer to actual AEC.","tokens_in":7488,"tokens_out":2813,"would_cite":false,"duration_ms":24633,"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":"A filter-bank neural network estimates echo delay with 91.67% accuracy and lifts mean speech quality by 0.633 PESQ.","keywords":["time delay estimation","acoustic echo cancellation","adaptive filter bank","residual echo suppression","OMLSA","automatic gain control","neural network classification","speech enhancement"],"falsifier":"Train or test the same TDE classifier with room impulse responses that are not pre-aligned to sample zero, or run it on recordings from a real room where the true echo delay is known by measurement, and compare the estimated delay against ground truth; if accuracy falls back to the correlation-baseline level, the sample-zero alignment is what carries the result.","tokens_in":6361,"feed_emoji":"🎙️","tokens_out":5790,"duration_ms":50218,"temperature":0.7,"pith_summary":"Spatial (acoustic) echo cancellation needs to know how far the far-end reference lags the microphone signal; when that delay is wrong, the adaptive filter wastes its length correcting the error instead of modeling the room. This paper proposes estimating the delay by running a bank of overlapping adaptive filters and classifying the pattern of their block-weight energies with a small neural network. On synthetic test data the estimator reaches 91.67% accuracy within a ±25 ms tolerance and 89.88% within ±5 ms, far above the open-source correlation-based AEC baseline. Around this estimator the paper builds a full echo-cancellation and speech-enhancement chain whose mean PESQ gain of 0.633 beats both the open-source AEC baseline and a much larger end-to-end neural network baseline.","feed_headline":"Echo-delay network hits 91.7% accuracy in tests","feed_subtitle":"Full echo-cancellation chain lifts mean speech quality by 0.633, beating a larger end-to-end model.","key_machinery":"The load-bearing mechanism is the adaptive filter bank. Five multidelay block frequency-domain adaptive filters, each split into $N=32$ blocks with $L=8$ blocks of overlap between adjacent filters, are fed from a shared buffer of the far-end signal; the $M\\times N$ block-energy values from all filters are concatenated into the input vector of a classification network taken from the voice-activity-detection path of a hybrid DSP/neural speech enhancer. The possible delays correspond one-to-one to classes, and the index of the maximal output probability is the estimated delay. This design converts a continuous cross-correlation search into a discrete pattern-recognition problem whose sensitivity to nonlinear echo and noise is governed by the learned classifier rather than by hand-tuned postprocessing.","core_discovery":"The paper claims that time delay estimation for acoustic echo cancellation can be recast as a classification problem: place a set of multidelay block frequency-domain adaptive filters so their time scopes overlap, compute the energy of each filter's weight blocks, concatenate those energies, and let a lightweight feed-forward classification network output the most probable delay. Because the filter bank itself performs linear echo estimation, its weight pattern already encodes where the echo sits; the classifier only has to read that pattern. The same best filter output then feeds a neural residual-echo suppressor followed by OMLSA-based denoising that treats residual echo as a special noise, and a sigmoid-smoothed automatic gain control. The paper reports that this combined system improves mean perceptual speech quality by 0.633 PESQ over the unprocessed signal, with a model size of 794K parameters, better than a 1827K-parameter end-to-end neural baseline and the open-source AEC baseline.","pith_inferences":["Editorial inference: because every RIR was pre-aligned with its direct path at sample zero, the reported TDE accuracy is probably an upper bound for real rooms, where the direct-path delay is unknown and must itself be discovered.","Editorial inference: the block-energy pattern of an adaptive filter bank may be a general, low-cost feature for any task that needs to localize a delayed copy of a reference signal, not just echo cancellation.","Editorial inference: one could test transfer directly by training the classifier on unaligned RIRs and measuring how much accuracy drops; that drop quantifies how much of the reported performance comes from the alignment shortcut."],"forward_implications":["A correlation-free TDE that loses only 1.79 percentage points when the tolerance tightens from $\\pm 25$ ms to $\\pm 5$ ms can support high-precision adaptive filtering without extra filter length.","The full AEC chain, at 794K parameters, outperforms an end-to-end neural AEC model more than twice its size in mean PESQ gain, so the linear-filter-first architecture is parameter-efficient.","Treating residual echo as a special noise within OMLSA reduces spectral distortion, lowering the variance of PESQ gain compared to neural suppression alone.","The sigmoid gain-smoothing rule lets the automatic gain control raise gain quickly without introducing frame-boundary discontinuities in the spectrum."],"supporting_citations":[{"why":"Supplies the multidelay block frequency-domain adaptive filter that forms each element of the filter bank.","marker":"[2]"},{"why":"Provides the end-to-end neural AEC baseline and the dataset preparation pipeline that the paper follows.","marker":"[14]"},{"why":"Supplies the OMLSA noise-estimation and suppression procedure that treats residual echo as a special noise.","marker":"[15]"},{"why":"Provides the lightweight classification network architecture (the voice-activity-detection path) used to map concatenated block energies to a delay class.","marker":"[16]"},{"why":"Supplies the clean speech and noise corpora used to generate the training data.","marker":"[17]"},{"why":"Supplies the room impulse responses used to synthesize near-end and echo signals in training.","marker":"[18]"},{"why":"Describes the procedure of identifying and setting the direct path of each room impulse response to position zero.","marker":"[19]"}],"fun_headline_variants":["Adaptive filter bank + neural net: 0.633 PESQ gain","Neural delay classifier boosts speech quality by 0.633","Echo-cancel via filter bank and neural classifier","Filter-bank neural approach improves echo cancel and speech","91.7% delay accuracy, 0.633 PESQ lift in one system"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training data assumes every room impulse response has its direct path shifted to sample zero, so the only delay the system must learn is the one artificially added during synthesis; real acoustic echo paths have unknown and varying direct-path delays that this training does not cover.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive filter bank + neural net: 0.633 PESQ gain","Neural delay classifier boosts speech quality by 0.633","Echo-cancel via filter bank and neural classifier","Filter-bank neural approach improves echo cancel and speech","91.7% delay accuracy, 0.633 PESQ lift in one system"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000297,"raw_usage":{"total_tokens":1695,"prompt_tokens":891,"completion_tokens":804,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":712}},"tokens_in":507,"tokens_out":804,"duration_ms":7458,"temperature":1.0,"reasoning_tokens":712,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:45:55.786788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train or test the same TDE classifier with room impulse responses that are not pre-aligned to sample zero, or run it on recordings from a real room where the true echo delay is known by measurement, and compare the estimated delay against ground truth; if accuracy falls back to the correlation-baseline level, the sample-zero alignment is what carries the result.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multidelay block frequency-domain adaptive filter that forms each element of the filter bank."},{"cited_title":"Cad-aec: Context-aware deep acoustic echo cancel- lation,","cited_arxiv_id":null,"evidence_quote":"Provides the end-to-end neural AEC baseline and the dataset preparation pipeline that the paper follows."},{"cited_title":"A neural network-based nonlinear acous- tic echo canceller,","cited_arxiv_id":null,"evidence_quote":"Supplies the OMLSA noise-estimation and suppression procedure that treats residual echo as a special noise."},{"cited_title":"Deep learning for acous- tic echo cancellation in noisy and double-talk scenar- ios,","cited_arxiv_id":null,"evidence_quote":"Provides the lightweight classification network architecture (the voice-activity-detection path) used to map concatenated block energies to a delay class."},{"cited_title":"Deep multitask acoustic echo cancellation.,","cited_arxiv_id":null,"evidence_quote":"Supplies the clean speech and noise corpora used to generate the training data."},{"cited_title":"Acoustic echo cancellation with the dual-signal transformation lstm network,","cited_arxiv_id":null,"evidence_quote":"Supplies the room impulse responses used to synthesize near-end and echo signals in training."},{"cited_title":"Speech enhance- ment for non-stationary noise environments,","cited_arxiv_id":null,"evidence_quote":"Describes the procedure of identifying and setting the direct path of each room impulse response to position zero."}],"review_version":1}