{"id":"f392b72c-43fb-4a5f-93e5-a40387ee0612","arxiv_id":"2603.02794","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A neural controller that predicts time-varying biquad filter coefficients can denoise speech with an interpretable equalizer chain, though the abstract and body report inconsistent specifications.","lead":"Researchers built a small neural network that controls 35 adjustable audio filters in real time to remove noise from speech, keeping the processing transparent instead of a black box. The paper claims this compact design roughly matches a much larger deep-learning denoiser on some quality metrics, but the abstract and full text disagree on model size and latency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's 24k-parameter, 10.7ms, HASPI/HASQI-verified model does not appear in the full text, making the headline result empirically unsubstantiated.","rationale":"The reader's REJECT is based on the internal inconsistency between the abstract and full text, and the underlying on-device claim being unsupported. I agree this is the most load-bearing concern: the paper's headline contribution is precisely the compact, hearing-aid-ready model, yet none of its headline numbers (24k params, 10.7ms, HASPI/HASQI, 29x MACs) are derivable from the experiments. The full-text model is larger and slower, and no hardware benchmark is provided. This is not a matter of disagreement with consensus but a direct falsifiability issue. A concrete check—reproducing the abstract's configuration and measuring the claimed hearing-aid metrics—would settle whether the central claim lands. Until then, the paper cannot be accepted as supporting its own abstract. The reader's weakest assumption correctly identifies the device-feasibility premise, and my concern is essentially the same one, sharpened by the metric/parameter mismatch.","tokens_in":9388,"tokens_out":5118,"duration_ms":45478,"concrete_test":"Request the 24k-parameter configuration (code/checkpoint) implied by the abstract and run it on the Valentini-Botinhao test set, computing HASPI/HASQI, MACs, and per-frame latency at 48 kHz. If the configuration does not exist, or if its HASPI/HASQI gap to DFNet3 exceeds about 0.02, the abstract's central claim must be withdrawn or revised to match the 1.01M-parameter, 21ms system actually evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim as advertised—a 24k-parameter, 10.7ms-latency, on-device model that stays within 0.02 of DFNet3 on hearing-aid metrics (HASPI/HASQI) with 29x fewer MACs—is absent from the experimental section. The full text reports a 1.01M-parameter model (§3.1), a 21ms algorithmic latency (§3.3), and a metrics list (§4.3) that contains no HASPI/HASQI. Table 1 reports no 24k row, no MAC counts, and no hearing-aid metrics. The abstract's inference-time controllability by mixing noisy and denoised inputs is also not described in the methodology. Thus the abstract and full text describe different systems, and the headline result cannot be checked against the reported data. Even the body's more modest claim of 'comparable' perceptual performance is selective: on eSTOI, LSD, and SI-SDR, DFNet3 outperforms TVF in Table 1. The on-device, real-time, hearing-aid-ready premise is therefore unsupported both by missing hardware measurements and by the numerical inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents TVF (Time-Varying Filtering), a speech denoising system in which a neural network predicts, per frame, the coefficients of a cascade of 35 differentiable biquad IIR filters. The model is trained on Valentini-Botinhao and compared against a static differentiable-PEQ baseline and a from-scratch DeepFilterNet3 baseline. The body reports a 1.01M-parameter model with 21 ms algorithmic latency, evaluated with PESQ, POLQA, eSTOI, LSD, SI-SDR, and SIGMOS. The abstract, however, advertises a 24k-parameter model with 10.7 ms latency, HASPI/HASQI hearing-aid metrics, and 29x fewer MACs than DFNet3, none of which appear in the full text.","tokens_in":9718,"tokens_out":3792,"duration_ms":37523,"significance":"The core idea—using a compact neural controller to drive an interpretable, time-varying IIR filter cascade—is timely and potentially useful for on-device speech enhancement. The systolic vectorization for training (§3.3) is a practical contribution, and the decision to retrain DFNet3 from scratch on the same data is methodologically fair. If the abstract's claims were supported, this would be a notable result: an interpretable and controllable model competitive with a much larger black-box model on perceptual metrics. As it stands, the evidence in the body supports only a much weaker statement: a 1.01M-parameter TVF is competitive on some metrics and worse on others, with no significance testing and no hardware validation.","major_comments":[{"comment":"The abstract's headline numbers are absent from the full text. The abstract claims 24k parameters, 10.7 ms latency, HASPI/HASQI results within 0.02 of DFNet3, and 29x fewer MACs, but §3.1 reports a 1.01M-parameter model, §3.3 gives 21 ms algorithmic latency, §4.3 lists no HASPI/HASQI among the metrics, and Table 1 reports no 24k row, no MAC counts, and no hearing-aid metrics. Additionally, the abstract calls DFNet3 \"almost two orders of magnitude larger\" than TVF, while the actual parameter ratio is 2.31M/1.01M ≈ 2.3x. The reader cannot verify the paper's central advertised result because the evaluated system appears to be different from the one described in the abstract.","section":"Abstract vs. §3.1, §3.3, §4.3, Table 1"},{"comment":"The claim that TVF is \"comparable\" to DFNet3 is selectively supported. On eSTOI (0.79 vs. 0.80), LSD (1.97 vs. 1.40), SI-SDR (13.71 vs. 14.58), and MOS-Signal (2.95 vs. 3.02), DFNet3 is numerically better; only on PESQ, POLQA, MOS-Noise, and MOS-Overall is TVF better. The standard deviations overlap substantially on all metrics, so without paired significance tests the observed differences are not established. The text's wording in §5.1, \"performs even better than DFNet3 on PESQ and POLQA,\" overstates the case given the lack of statistical testing and the mixed pattern across metrics.","section":"§5.1, Table 1"},{"comment":"The abstract promises an inference-time controllability mechanism: \"it can be set during training through the loss weighting, and adjusted at inference, with no retraining, by mixing the noisy input with the denoised output.\" The full text, however, only describes adjusting the time-domain loss weight during training (§5.2). The inference-time mixing operation is never defined in the methodology and no experiment demonstrates it. This is a missing component of the claimed contribution, not merely a presentation issue.","section":"Abstract vs. §5.2"},{"comment":"The on-device, hearing-aid-ready premise is unsupported. The paper provides an algorithmic latency of 21 ms (§3.3) but no actual runtime, power, or memory measurements on any device. The abstract's 10.7 ms and 29x MAC figures do not appear in the body. The claim that the model is \"within hearing-aid budgets\" is therefore an assertion without evidence, and the title's \"On-Device Assistive Hearing\" is not validated by any hardware experiment.","section":"§3.3, Abstract"},{"comment":"The assertions of \"artifact-free linear processing\" (§6) and that the GRU smoothly changes coefficients \"avoid[ing] artifacts\" (§5.3) are presented as conclusions without any artifact-specific evaluation or listening test. Figure 2 is a single qualitative example; it cannot substantiate a general artifact-free claim. This is a strong claim that requires either a perceptual listening test, an artifact metric, or a clearly hedged statement.","section":"§5.3, §6"}],"minor_comments":[{"comment":"Reference [4] contains a typo: \"V ocoder\" for \"Vocoder\"; reference [5] has \"V ocoder\" as well. Please proofread the reference list.","section":"§2, References"},{"comment":"The author names include \"ID\" strings and a double asterisk that is not explained in the footnote. These appear to be artifacts of the submission template and should be cleaned.","section":"Author block"},{"comment":"The spectrograms in Figure 2 would benefit from time/frequency axis labels and display ranges; without them it is hard to assess the claimed -40 dB attenuation and 0 dB passband behavior.","section":"Figure 2"},{"comment":"The DFNet3 retraining uses \"default hyperparameters\" from the official repository, but the paper does not specify whether the optimizer schedule, loss weights, or data sampling were tuned for the much smaller Valentini-Botinhao dataset. A sentence acknowledging this limitation would strengthen the fairness discussion.","section":"§4.2"},{"comment":"The MOS-Noise/MOS-Signal trade-off is described as \"tunable\" based on the loss weight, but no experiment varies this weight. Consider adding an ablation or explicitly labeling this as future work.","section":"§5.2"}],"recommendation":"reject","confidential_remarks":"The discrepancy between the abstract and the full text is severe: the abstract describes a different model (24k parameters, 10.7 ms, HASPI/HASQI) and different claims than the evaluated system. Even setting aside the abstract, the body's main comparative claim is weakened by mixed metric results and missing significance tests. These are load-bearing issues that cannot be fixed by local revision; the paper would need new experiments and a new abstract. I therefore recommend rejection, though I would be willing to consider a substantially revised version that honestly reports the 1.01M-parameter model and validates or removes the on-device/hearing-aid claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The abstract and the body describe different systems. The abstract's headline — 24k parameters, 10.7 ms latency, HASPI/HASQI within 0.02 of DFNet3, 29x fewer MACs, and inference-time controllability — does not appear in the experimental section. The body reports a 1.01M-parameter model (§3.1), 21 ms algorithmic latency (§3.3), and evaluation with PESQ, POLQA, eSTOI, SI-SDR, LSD, and SIGMOS (§4.3). No HASPI/HASQI, no MAC counts, no 24k row in Table 1, and the inference-time input mixing claimed in the abstract is not described in the methodology. That is a load-bearing inconsistency, not a cosmetic slip.\n\nThat said, the underlying idea is genuinely new. A causal, neural-controlled biquad cascade for real-time speech denoising is a legitimate application — related dynamic EQ [11] is non-causal, time-varying biquads [9] target offline processing, and [10] uses FIR in a feedback loop. The systolic vectorization for training is a useful trick. The body also does some things right: all models are trained from scratch on the same data, DFNet3's lower-than-original numbers are acknowledged, and the conclusion explicitly calls it a proof of concept. TVF does beat the static PEQ baseline on essentially all metrics, which supports the value of time-varying control. On the DFNet3 comparison, it wins on PESQ, POLQA, MOS-Noise, and MOS-Overall, but loses on SI-SDR and LSD, with eSTOI about tied. That is \"comparable\" in a loose sense, but there are no significance tests, and SIGMOS is a learned proxy, not a listening test. For a hearing-aid claim, the absence of any hardware measurement — power, on-device latency, actual HA processing — is a real gap. The 1M-parameter GRU isn't obviously \"ultra-lightweight\" for a hearing aid.\n\nThe soft spots are in proportion: the central idea holds up as a proof of concept, but the submitted paper oversells it. The mismatch alone means the headline result cannot be checked against the reported data. Still, the body is worth reading for anyone working on interpretable DSP/ML denoisers.\n\nWho is this for: people in hybrid DSP/ML speech enhancement; it would get a fair hearing at a workshop or INTERSPEECH-style venue after the authors fix the abstract, add MAC/latency numbers, and ideally a listening test or hardware profile. My recommendation: don't accept as is, but don't dismiss the work. If I were the editor, I'd ask the authors to reconcile the abstract and body and supply the missing evidence, then send it to a serious referee. It deserves a round of review, not a silent desk rejection.","headline":"The abstract and the body describe different models — the headline 24k-parameter, 10.7ms, HASPI/HASQI result is absent from the full text — but the causal neural biquad-cascade idea is new, and the body's results, though selective, support a more modest proof-of-concept claim.","tokens_in":10187,"tokens_out":3780,"would_cite":false,"duration_ms":35704,"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 1M-parameter neural controller that sets 35 biquad filters frame by frame matches a 2.3M-parameter black-box denoiser on perceptual scores.","keywords":["speech enhancement","time-varying IIR filters","biquad cascade","interpretable DSP","on-device inference","differentiable filtering","noise suppression","low latency"],"falsifier":"Measure end-to-end inference time and energy on a low-power DSP or hearing-aid chip using the released model; if the 1.01M-parameter GRU cannot run within a few milliseconds and a few milliwatts, the on-device claim fails. Alternatively, retrain the 24k-parameter version and check whether it still stays within 0.02 of the deep baseline on HASPI/HASQI.","tokens_in":9318,"feed_emoji":"🎧","tokens_out":4373,"duration_ms":38961,"temperature":0.7,"pith_summary":"The paper claims that a compact neural network can control a cascade of 35 second-order IIR filters (biquads) in real time, predicting gain, center frequency, and Q per frame from the magnitude spectrum. This time-varying filtering chain tracks non-stationary noise while keeping every spectral modification explicit as an adjustable equalizer curve. Trained from scratch on a standard noisy-speech corpus, the 1.01M-parameter model matches a state-of-the-art deep denoiser with 2.31M parameters on PESQ, POLQA, and overall MOS, while using a strictly linear, causal signal path. The authors present this as a proof of concept for interpretable, on-device assistive hearing.","feed_headline":"A small neural net steers 35 biquad filters to rival a bigger denoiser","feed_subtitle":"Interpretable time-varying filtering matches a 2.3M-parameter black box on perceptual scores, with a fully auditable equalizer path.","key_machinery":"The key object is the differentiable cascade of 35 biquad IIR filters, parameterized by gain, quality factor, and center frequency and updated every frame. A gated recurrent unit generates smooth coefficient trajectories to avoid audible discontinuities, and a systolic-style vectorized tensor formulation accelerates training; at inference the cascade runs serially with 21ms algorithmic latency.","core_discovery":"The central claim is that a GRU-based controller with about 1M parameters can map a 513-bin magnitude spectrum to 105 biquad coefficients (35 filters × 3 parameters) per 21ms frame, and that the resulting time-varying linear filter both adapts to changing noise and remains fully interpretable. In head-to-head training on the same corpus, TVF outperforms a static 35-band equalizer and roughly ties or beats a deep STFT-mask baseline on perceptual metrics, though it trails on waveform-reconstruction metrics like SI-SDR and log-spectral distance. The abstract asserts a much smaller 24k-parameter variant at 10.7ms latency, but the full text and experiments use the 1.01M-parameter model.","pith_inferences":["If the claimed latency and parameter counts hold on real hardware, the same controller-plus-biquad architecture could be retargeted to other audio effects (EQ, compression, feedback suppression) with minimal changes.","The paper's abstract claims a 24k-parameter model at 10.7ms latency, while the body describes 1.01M parameters and 21ms; reconciling this on actual silicon would be the decisive next test.","The linear, causal constraint suggests a natural ceiling: in very low SNR or with transient interference, a phase-reconstructing nonlinear mask may still be unavoidable.","One could test the interpretability claim directly by asking hearing-impaired listeners to adjust the exposed filter curves to their preference."],"forward_implications":["A denoiser whose entire signal path is a linear equalizer can be examined and audited: every gain change corresponds to a visible filter curve.","The suppression-versus-preservation tradeoff can be rebalanced at inference without retraining by mixing noisy and denoised audio, enabling user-controlled hearing-aid behavior.","Because the model is structurally causal and low-latency, it can run inside a hearing aid's pipeline rather than on a server.","On small training corpora, the inductive bias of linear IIR filtering yields competitive perceptual quality with far fewer parameters than black-box masks."],"fun_headline_variants":["Interpretable denoiser approaches a 2.3M-param black box on perceptual scores","Tiny neural net tunes 35 biquads for auditable noise removal","Time-varying biquad filters demystify speech denoising","On-device hearing aid uses interpretable equalizer to approach big denoisers","Lightweight controller makes biquad denoiser interpretable"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The real-time on-device claim rests on an untested hardware assumption: the paper states 1.01M parameters and 21ms latency in the body but the abstract promises 24k parameters and 10.7ms, with no power or runtime measurement on a hearing-aid-class processor.","fun_headline_variants_meta":{"raw":{"variants":["Interpretable denoiser approaches a 2.3M-param black box on perceptual scores","Tiny neural net tunes 35 biquads for auditable noise removal","Time-varying biquad filters demystify speech denoising","On-device hearing aid uses interpretable equalizer to approach big denoisers","Lightweight controller makes biquad denoiser interpretable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001314,"raw_usage":{"total_tokens":5241,"prompt_tokens":845,"completion_tokens":4396,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":4309}},"tokens_in":589,"tokens_out":4396,"duration_ms":30279,"temperature":1.0,"reasoning_tokens":4309,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T19:14:22.352236+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure end-to-end inference time and energy on a low-power DSP or hearing-aid chip using the released model; if the 1.01M-parameter GRU cannot run within a few milliseconds and a few milliwatts, the on-device claim fails. Alternatively, retrain the 24k-parameter version and check whether it still stays within 0.02 of the deep baseline on HASPI/HASQI.","supporting_citations":[],"review_version":1}