{"id":"e3fae235-e832-497f-a831-4696e25a0b8b","arxiv_id":"2502.03260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An adaptive Gabor filterbank front-end controlled by a neural feedback loop outperforms the LEAF learnable front-end on eight speech, sound event, and music classification tasks.","lead":"The paper compares an adaptive audio front-end that adjusts its filters in real time against standard learnable front-ends across eight audio benchmarks. It finds the adaptive front-end consistently more accurate and more stable across training epochs, suggesting input-adaptive feature extraction is worth pursuing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central robustness claim is inferred from flat test-accuracy curves on clean recordings; no experiment varies acoustic conditions, so the adaptive-front-end thesis is not yet supported.","rationale":"The paper's stated purpose is to determine whether audio front-ends should be adaptive, with the operative distinction being that non-adaptive front-ends are 'fixed computation graphs at inference time, failing to dynamically adapt to varying acoustic environments.' The headline empirical contribution is therefore not merely that Ada-FE obtains higher clean accuracy, but that adaptivity yields 'stability or robustness.' The only evidence for this is the flatness of test learning curves in Figures 11-14. Learning-curve flatness is a property of the optimization trajectory under one training setup; it does not measure response to acoustic variation, and no test-time perturbation is applied anywhere in the study. This makes the inference from curve flatness to acoustic robustness the weakest load-bearing step in the argument. The accuracy comparisons in Table III are better supported: Table II includes a no-adaptation ablation, several accuracy gaps are large, and the LEAF baseline was run with the official code under the same settings. However, those clean-data accuracy numbers alone do not justify the adaptive-front-end thesis as framed in the abstract and conclusion. The reader's weakest_assumption identifies exactly this gap, and I agree with it. Other issues, such as missing error bars, TD-fbanks numbers copied from the literature, and the contradictory statement about LDA in the conclusion, are real but secondary; they do not bear as directly on the central conceptual claim. The verdict should remain conditional: the paper should not be rejected, but the robustness claim needs a direct acoustic-perturbation test before it can be accepted as the paper's central contribution.","tokens_in":19676,"tokens_out":6075,"duration_ms":58937,"concrete_test":"Retrain or reuse the Ada-FE and LEAF checkpoints with the EfficientNet-B0 back-end and evaluate them on held-out test audio from SPC-V2 or VoxCeleb1 under (a) clean conditions, (b) additive noise (e.g., MUSAN babble at 0, 10, and 20 dB SNR), and (c) reverberation (e.g., convolution with measured room impulse responses). Report Top-1 accuracy for each condition. If Ada-FE's gain over LEAF under corrupted conditions is no larger than on clean audio, the flat-curve inference is invalid and the 'adapt to varying acoustic environments' claim lacks direct support. If the corrupted-condition gains are consistently larger, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper motivates Ada-FE by the claim that fixed/learnable front-ends 'fail to dynamically adapt to varying acoustic environments' (Abstract) and concludes that Ada-FE 'shows significantly better stability than the baseline on the test set, over various epochs' (Section VI). The only evidence offered is in Section V.C and Figures 11-14: test-accuracy learning curves that are flatter for Ada-FE and Ada-FE-S-FM than for LEAF. The text then says 'The learning curves of LEAF suggest that LEAF is quite sensitive to varying acoustic conditions' (Section V.C). This is an unsupported inference. All eight benchmarks use clean recordings: no noise, reverberation, or channel distortion is applied at train or test time, and no domain-shift evaluation is reported. Flatness of accuracy versus training epoch measures optimization stability of one training run, not invariance of the learned representation to acoustic environment. Because the abstract's headline advantage ('stability or robustness') and the conclusion's 'advantages of employing feedback control ... to respond to varying conditions' rest on this inference, the central claim that adaptive front-ends are superior because they adapt is not established even if the raw clean-data accuracy numbers in Table III are correct. The raw accuracy claim is a separate, better-supported empirical claim; the adaptive robustness claim is the load-bearing one and is currently underwritten only by an analogy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper asks whether audio front-ends should be input-adaptive, and evaluates Ada-FE, a Gabor-filter front-end with a neural feedback controller that adjusts Q-factors frame-wise, against LEAF and TD-fbanks on eight audio benchmarks with two back-ends. It also introduces Ada-FE-S, a simplified version with the hand-crafted LDA module removed, and ablates the controller input choices. The main empirical claims are that Ada-FE and Ada-FE-S-FM outperform LEAF and TD-fbanks across all tasks, and that Ada-FE exhibits better stability or robustness on test samples over training epochs.","tokens_in":19944,"tokens_out":6891,"duration_ms":52092,"significance":"If the accuracy advantage is reproducible, the paper provides useful evidence that input-adaptive front-ends can outperform fixed learnable front-ends, and the 'w/o Adaptation' ablation is a valuable control. The evaluation across two backbones and eight tasks, the use of the official LEAF implementation, and the repeated-runs protocol are strengths. However, the central robustness claim is not supported by the clean-data experiments, and several comparison details (TD-fbanks coverage, error bars, baseline specification) need to be tightened before the headline claims can be accepted.","major_comments":[{"comment":"The stability/robustness claim is not supported by the reported experiments. Figures 11-14 show test-accuracy learning curves on clean recordings, and no experiment applies noise, reverberation, channel distortion, or a domain shift at test time. The statement in Section V.C that 'The learning curves of LEAF suggest that LEAF is quite sensitive to varying acoustic conditions' is an inference from optimization dynamics on matched clean data, not from any controlled acoustic variation. Because the abstract's 'stability or robustness' claim and the conclusion's 'respond to varying conditions' rest on this inference, the adaptive-robustness claim is under-supported. Either add experiments with acoustic degradations or revise the claim to 'stable test accuracy across epochs'.","section":"Section V.C, Section VI, Abstract"},{"comment":"The claim that Ada-FE and Ada-FE-S-FM 'outperform LEAF and TD-fbanks, across all the tasks' is not supported by Table III. TD-fbanks results are reported only for SPC-V2 (87.70) and VoxCeleb1 (26.0) under EfficientNet-B0, and these are taken from the original LEAF paper [19]; no TD-fbanks numbers are given for ESC-50, GTZAN, FMA-S, CREMA-D, IEMOCAP, SPC-V1, or for the MobileNetV2 backbone. Please either re-run TD-fbanks under the same protocol on all benchmarks or restrict the superiority claim to LEAF.","section":"Section V.C, Table III"},{"comment":"The paper reports means over three runs but no error bars, confidence intervals, or significance tests are provided. Many head-to-head differences are small (e.g., Ada-FE vs Ada-FE-S-FM on SPC-V1 with EfficientNet-B0: 91.29 vs 91.14 in Table II), so the reader cannot tell whether the reported ordering is reliable. Please report per-run variation or statistical significance for the main comparisons, at least for the Ada-FE vs LEAF differences.","section":"Section IV.B, Section V.C"},{"comment":"The conclusion that removing LDA demonstrates 'hand-crafted LDA is quite essential' contradicts the body of the paper. In Section V.B and Table II, Ada-FE-S-FM (without LDA) achieves comparable or better accuracy than Ada-FE on most benchmarks, and the text states that the Q-factor adaptive learning 'benefits from removing the hand-crafted module.' The conclusion should be corrected.","section":"Section VI, Section V.B"},{"comment":"The 'w/o Adaptation (Not Learnable)' baseline is the key control for the adaptivity claim, but its configuration is not specified in the text. It is unclear whether the Gabor filters are fixed at their initial Q values, at some learned static Q values, or with only the fixed Gabor layer active. Please define the exact front-end used in this condition so that the adaptivity benefit can be properly interpreted.","section":"Table II, Section V.B"}],"minor_comments":[{"comment":"The word 'netwroks' should be 'networks'.","section":"Table III caption"},{"comment":"The model name 'Ada-FE-S-F' should be 'Ada-FE-S-FM'.","section":"Table IV header"},{"comment":"The dataset name 'V oxCeleb1' should be 'VoxCeleb1' without the stray space.","section":"Throughout"},{"comment":"The word 'MobilenNetV2' should be 'MobileNetV2'.","section":"Figure 12 caption"},{"comment":"The reference to 'Jailty and Hinton' should be 'Jaitly and Hinton'.","section":"Section II"},{"comment":"The term 'Gammaton-filters' should be 'Gammatone-filters'.","section":"Section II"},{"comment":"The phrase 'repost the accuracy' should be 'report the accuracy'.","section":"Section IV.A"},{"comment":"The typo 'TD-fbansks' should be 'TD-fbanks'.","section":"Section IV.B"}],"recommendation":"major_revision","confidential_remarks":"The heavy self-citation is noticeable, but the core comparisons use the official LEAF code and public datasets, so I do not see a circularity problem. The robustness overclaim is the main risk; the authors should be asked to run at least one noise/domain-shift experiment or substantially soften the claim. The editorial decision should hinge on whether the accuracy advantage survives with proper statistical and baseline coverage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the clean-data accuracy comparison is probably the most complete head-to-head of an adaptive front-end against LEAF that exists, and the ablation showing the adaptive controller matters is real evidence. But the headline 'stability or robustness' claim is not supported by the experiments they ran, and the conclusion contains a contradiction about the LDA module that needs fixing.\n\nWhat is actually new: they take their own Ada-FE (previously published for replay spoofing detection) and evaluate it on eight audio classification benchmarks with two back-ends, and they introduce a simplified Ada-FE-S that removes the hand-crafted level-dependent adaptation. The simplification result is a genuine design contribution: Ada-FE-S-FM matches or beats the full model on most tasks, suggesting the neural feedback controller alone can do the work. The comparisons use official LEAF code and equal back-end architectures, which is the right way to run this.\n\nWhere it gets shaky: the stability/robustness claim. The evidence is Figures 11-14, which show test-accuracy learning curves over training epochs that are flatter for Ada-FE than for LEAF. That is optimization stability of one training run, not invariance to acoustic environment. All eight datasets are clean recordings; no noise, reverberation, or channel distortion is applied at train or test time, and no domain-shift evaluation appears. The abstract motivates adaptive front-ends by saying fixed ones 'fail to dynamically adapt to varying acoustic environments,' and the conclusion says Ada-FE responds to 'varying conditions.' The experiments never vary the acoustic environment. The stress-test note is right: that part of the thesis is unsupported. The clean-data accuracy advantage over LEAF is a separate, better-supported claim.\n\nOther soft spots: no error bars or significance tests despite averaging three runs; on some datasets the gap is small, so a few runs could change the picture. TD-fbanks numbers are taken from the LEAF paper rather than rerun, which is common but weakens the comparison. No code release. And the LDA contradiction is concrete: Section V.B says removing LDA is beneficial, while the conclusion says 'demonstrating hand-crafted LDA is quite essential.' Presumably that should be 'not essential,' but as written it confuses the paper's own summary.\n\nBottom line: this paper is worth refereeing. The benchmark breadth and the design simplification justify expert time, and the central accuracy claim is plausible. A good reviewer will ask for either noise/domain-shift experiments or a rewrite that scopes the robustness claim to training stability. I would cite the clean-data results, not the robustness claim.","headline":"Solid clean-data comparison and a useful simplification, but the robustness claim rests on a misinterpreted learning-curve plot.","tokens_in":20491,"tokens_out":3051,"would_cite":true,"duration_ms":25707,"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":"Adaptive audio front-ends beat fixed learnable ones on eight audio tasks.","keywords":["Representation learning","Audio front-end","Adaptive inference","Gabor filters","Learnable front-ends","Audio classification","Neural feedback controller","Q-factor"],"falsifier":"Corrupt the test portions of ESC-50, Speech Commands V2, and VoxCeleb1 with additive noise, reverberation, or band-limiting at several signal-to-noise ratios and compare Ada-FE against LEAF; if LEAF matches or beats Ada-FE under those acoustic degradations, the claim that adaptive inference helps in varying acoustic environments would fail even if the clean-data accuracy advantage stands.","tokens_in":19480,"feed_emoji":"🎧","tokens_out":7465,"duration_ms":58181,"temperature":0.7,"pith_summary":"This paper asks whether the front-end of an audio classifier—the stage that converts raw waveforms into features—should adapt to each input at inference time instead of staying fixed after training. To answer it, the authors compare Ada-FE, a front-end whose Gabor filters are reshaped frame by frame by a neural feedback controller, with LEAF and TD-fbanks on eight speech, sound-event, and music benchmarks using two back-end classifiers. They report that the adaptive front-ends are more accurate than the fixed learnable front-ends on every task and with both back-ends, in some cases by large margins, and that their test accuracy is more stable across training epochs. A simplified version that drops the hand-crafted level-dependent adaptation matches or exceeds the full model, suggesting the neural controller can learn the needed level-dependent behaviour on its own. The paper reads this as evidence that input-adaptive front-ends are a better default than fixed learnable front-ends across audio classification tasks.","feed_headline":"Adaptive audio front-ends beat fixed learnable ones on 8 tasks","feed_subtitle":"A feedback controller reshapes filters frame by frame, improving accuracy and training stability over LEAF.","key_machinery":"The central object is an adaptive Gabor filter layer whose only tunable shape parameter is the Q-factor, defined as the ratio of center frequency to bandwidth, $Q = f_c / BW$. The filter's frequency response at the center frequency satisfies $|W(\\Omega_c)| \\approx \\sqrt{2\\pi}\\, Q / \\Omega_c$, so raising $Q$ narrows the filter and raises its gain, while lowering $Q$ widens it and compresses the input. Each frame's Q value is the sum of two parallel terms: $Q_E$ from a hand-crafted level-dependent adaptation that reads subband energy (the amplitude-modulation pathway) and $Q_{FM}$ from a two-layer fully-connected neural adaptive feedback controller fed with frequency-modulation features. The entire system is trained end-to-end with a back-end classifier, with the controller's gradient path running counter to the main network path. A simplified version, Ada-FE-S, removes the hand-crafted $Q_E$ term so the neural controller alone determines the adaptive Q value.","core_discovery":"On the paper's own terms, the central discovery is that adapting filter shape during inference, not just learning it during training, improves audio classification. Ada-FE operates on raw waveforms with a fixed Gabor filter layer followed by an adaptive Gabor filter layer; the adaptive layer's Q-factor, defined as the ratio of center frequency to bandwidth, is updated every frame as the sum of a level-dependent term computed from subband energy by a hand-crafted piecewise function and a term from a neural adaptive feedback controller fed with frame-averaged frequency-modulation features. Because the Q-factor directly sets gain and bandwidth through $|W(\\Omega_c)| \\approx \\sqrt{2\\pi}\\, Q/\\Omega_c$, the controller can make a filter more selective or more compressive as the input changes. Across ESC-50, GTZAN, FMA-S, CREMA-D, IEMOCAP, Speech Commands V1/V2, and VoxCeleb1, Ada-FE and its simplified variant Ada-FE-S-FM outperform LEAF on Top-1 accuracy with both EfficientNet-B0 and MobileNetV2-100 back-ends, with relative gains up to about 34% on VoxCeleb1 with 1-second segments. The paper also reports that removing the hand-crafted level-dependent adaptation does not hurt, and that removing the fixed first filter layer does hurt, so the fixed stage remains structurally important. The conclusion is that a feedback-controlled, input-adaptive front-end offers advantages over a fixed learnable front-end.","pith_inferences":["Because the paper only tests clean recordings, the reported accuracy advantage over LEAF is a lower bound on the benefit of adaptivity: the strongest gains would be expected under mismatched acoustic conditions such as noise or reverberation, which a straightforward test-set corruption experiment could measure.","The frame-wise Q-factor trajectories Ada-FE produces encode a per-frame estimate of the acoustic environment; exposing these trajectories to the back-end, or using them as an auxiliary training signal, is a testable extension the paper does not explore.","Plugging Ada-FE or Ada-FE-S into self-supervised pre-training pipelines is a natural next step, and the paper itself names pre-training as future work; a failure to see gains there would bound the value of adaptivity to low-level feature extraction."],"forward_implications":["Ada-FE and Ada-FE-S-FM improve Top-1 accuracy over LEAF on all eight datasets with both EfficientNet-B0 and MobileNetV2-100 back-ends.","The simplified adaptive front-end Ada-FE-S-FM, which removes the hand-crafted level-dependent adaptation, matches or exceeds the full Ada-FE on most tasks.","The fixed first Gabor filter layer is still needed: removing it degrades accuracy substantially on every benchmark.","Adaptive front-ends reach high accuracy with far fewer training epochs than LEAF, and their test accuracy remains stable across epochs, so early stopping is less risky."],"supporting_citations":[{"why":"Supplies the main learnable front-end baseline (LEAF) that Ada-FE must outperform.","marker":"[19]"},{"why":"Defines the learnable time-domain filterbank (TD-fbanks) used as the second baseline.","marker":"[1]"},{"why":"Introduces the original Ada-FE adaptive front-end that this paper systematically evaluates and simplifies.","marker":"[26]"},{"why":"Provides the biologically inspired adaptive-Q filterbank and level-dependent adaptation mechanism.","marker":"[44]"},{"why":"Supplies the frame-averaged frequency-modulation input used by the neural adaptive feedback controller.","marker":"[45]"},{"why":"Defines the spectral-centroid-deviation method used to compute the FM features.","marker":"[48]"},{"why":"Provides the ESC-50 environmental sound classification benchmark.","marker":"[5]"},{"why":"Provides the Speech Commands V1/V2 keyword spotting datasets.","marker":"[56]"},{"why":"Provides the VoxCeleb1 speaker identification dataset.","marker":"[57]"}],"fun_headline_variants":["Adaptive audio front-end beats fixed learnable front-ends on 8 tasks","Adaptive filter Q-factors boost audio classification on 8 benchmarks","Inference-time filter adaptation outperforms learned static front-ends","Adaptive front-end retunes filters per frame to outperform LEAF","Adaptive audio front-end yields up to 34% relative gain over LEAF"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the flatness of test-accuracy learning curves over training epochs is evidence of robustness to varying acoustic environments, even though all evaluations use clean recordings with no added noise, reverberation, or channel distortion at test time.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive audio front-end beats fixed learnable front-ends on 8 tasks","Adaptive filter Q-factors boost audio classification on 8 benchmarks","Inference-time filter adaptation outperforms learned static front-ends","Adaptive front-end retunes filters per frame to outperform LEAF","Adaptive audio front-end yields up to 34% relative gain over LEAF"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000968,"raw_usage":{"total_tokens":4173,"prompt_tokens":1057,"completion_tokens":3116,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":673,"completion_tokens_details":{"reasoning_tokens":3020}},"tokens_in":673,"tokens_out":3116,"duration_ms":20640,"temperature":1.0,"reasoning_tokens":3020,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:21:52.066642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Corrupt the test portions of ESC-50, Speech Commands V2, and VoxCeleb1 with additive noise, reverberation, or band-limiting at several signal-to-noise ratios and compare Ada-FE against LEAF; if LEAF matches or beats Ada-FE under those acoustic degradations, the claim that adaptive inference helps in varying acoustic environments would fail even if the clean-data accuracy advantage stands.","supporting_citations":[{"cited_title":"LEAF: A Learnable Frontend for Audio Classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the main learnable front-end baseline (LEAF) that Ada-FE must outperform."},{"cited_title":"Learning filterbanks from raw speech for phone recogni- tion,","cited_arxiv_id":null,"evidence_quote":"Defines the learnable time-domain filterbank (TD-fbanks) used as the second baseline."},{"cited_title":"Dnn controlled adaptive front-end for replay attack detection systems,","cited_arxiv_id":null,"evidence_quote":"Introduces the original Ada-FE adaptive front-end that this paper systematically evaluates and simplifies."},{"cited_title":"Biologically in- spired adaptive-q filterbanks for replay spoofing attack detection,","cited_arxiv_id":null,"evidence_quote":"Provides the biologically inspired adaptive-Q filterbank and level-dependent adaptation mechanism."},{"cited_title":"Replay detection in voice biometrics: an investiga- tion of adaptive and non-adaptive front-ends,","cited_arxiv_id":null,"evidence_quote":"Supplies the frame-averaged frequency-modulation input used by the neural adaptive feedback controller."},{"cited_title":"Detection of replay-spoofing attacks using frequency modulation features,","cited_arxiv_id":null,"evidence_quote":"Defines the spectral-centroid-deviation method used to compute the FM features."},{"cited_title":"ESC: Dataset for environmental sound classification,","cited_arxiv_id":null,"evidence_quote":"Provides the ESC-50 environmental sound classification benchmark."},{"cited_title":"V oxceleb: A large-scale speaker identification dataset,","cited_arxiv_id":null,"evidence_quote":"Provides the VoxCeleb1 speaker identification dataset."}],"review_version":1}