{"id":"1d13cd17-7f1b-4143-ab49-e03ce8bb5197","arxiv_id":"2412.00312","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A CNN whose filters are learnable cosines (two parameters each) beats a comparable standard CNN on raw audio classification and, with added memory and vector quantization, posts strong results on five benchmarks.","lead":"The authors replace ordinary CNN filters with learnable cosine filters, each using only two parameters, and report better accuracy with roughly 77% fewer parameters on five audio classification datasets. They also add vector quantization and memory to build VQCCM, which they claim reaches state-of-the-art results, a claim their own table contradicts on one dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The protocol selects CosCovNN and VQCCM hyperparameters on S09/IEMOCAP using best-of-five accuracy and then reports final accuracy on those same datasets, so the claimed gains over CNN/LEAF may be selection artifacts.","rationale":"I focus on the selection protocol because it directly affects the head-to-head evidence for both CosCovNN-vs-CNN and VQCCM-vs-LEAF. The reader's weakest assumption identified the same issue; my read agrees and sharpens it by noting that the same S09 and IEMOCAP splits used for selection are then reported as final results. If a held-out validation split were used for selection, the reported margins might shrink. The SOTA contradiction is also real and worth fixing in revision, but it is an overstatement rather than a threat to the core architecture claim. For these reasons the conditional verdict is appropriate; I do not move to reject because the underlying idea, a two-parameter cosine filter, is plausible and the flaws are addressable with better experimental hygiene and corrected claims.","tokens_in":17430,"tokens_out":14696,"duration_ms":134114,"concrete_test":"Re-run the full architecture search of §5.1.1 on a held-out validation split of S09 (e.g., 80/10/10) using the mean or median accuracy over five seeds instead of the maximum, then evaluate the chosen CosCovNN and the equivalent CNN on the untouched test split. If the test accuracy of CosCovNN no longer exceeds the CNN baseline, or the S09/Spoken Digit entry drops materially below 96.3%, the claimed 77%-fewer-parameter advantage is at least partly a selection artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim rests on the accuracy comparisons in Table 4. Those comparisons are not trustworthy as reported because the model-selection protocol uses the maximum accuracy over five runs on the same datasets that later appear in the evaluation table. Section 5.1.1 states that the architecture search on S09 'is conducted five times and the maximum accuracy is recorded for comparison,' and the resulting filter sizes (100, 50, 12, 6, 3) and pooling sizes (10, 8, 4, 4) are the ones used in Table 4 for 'Spoken Digit' (S09). Section 5.4.1 similarly uses 'maximum accuracy based on five runs' on S09 and IEMOCAP to choose VQCCM memory and VQ sizes, and both datasets appear as final rows in Table 4. Selecting hyperparameters by best-of-five on the final evaluation data can inflate results; the reported mean±std for the chosen configuration does not correct for this selection bias. This is especially consequential for the small margins in Table 4 (e.g., VQCCM 97.1 vs LEAF 96.7 on Spoken Digit; 99.1 vs 99.1 on Acoustic Scenes). The abstract's SOTA claim is also internally contradicted by Table 4 (TD-fbanks 99.5 > VQCCM 99.1 on Acoustic Scenes), an additional sign that the numbers are being interpreted too generously.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces CosCovNN, a convolutional architecture for raw-waveform audio classification in which each filter is a learnable cosine function g[n] = θ₁ cos(θ₂ n), requiring only two parameters per filter instead of L. It also presents VQCCM, an extension that adds vector quantization and a memory layer. The authors report that CosCovNN outperforms an equivalent CNN with 77.66% fewer parameters on five audio datasets (Speech Command, Spoken Digit, Speech Emotion, Acoustic Scenes, Musical Instrument, Speaker Id), and that VQCCM achieves state-of-the-art accuracy on all five datasets. The supporting material includes an architecture search on the S09 dataset, comparisons to TD-fbanks, SincNet, LEAF, and a CNN baseline, and an ablation study of the memory and VQ components.","tokens_in":17833,"tokens_out":5558,"duration_ms":46728,"significance":"If the core claims hold, the cosine-filter parameterization is an elegant and practically attractive idea: it drastically reduces the number of trainable parameters in convolutional layers while preserving or improving accuracy on raw-waveform audio. The paper's central comparison is internally coherent in its parameter-count arithmetic (the 77.66% reduction follows from the filter sizes and the two-parameter form), and the use of a straightforward Fourier/DCT-inspired filter is a clear, easily communicated contribution. The authors also provide a concrete architecture and an ablation that isolates the effects of memory and VQ components. However, the experimental protocol and result reporting as presented do not currently support the paper's headline claims: the selection-on-the-evaluation-set protocol introduces selection bias, and the SOTA claim is contradicted by the paper's own Table 4 on Acoustic Scenes. The contribution is worth pursuing, but the empirical case needs to be rebuilt with proper validation methodology and corrected claims.","major_comments":[{"comment":"The model-selection protocol uses the maximum accuracy over five training runs on S09 and IEMOCAP, and those very datasets then appear in the final evaluation in Table 4. Section 5.1.1 states that the S09 architecture search 'is conducted five times and the maximum accuracy is recorded for comparison,' and Section 5.4.1 says that VQCCM memory and VQ sizes were chosen using 'maximum accuracy based on five runs' on S09 and IEMOCAP. The reported mean±std for the chosen configuration does not correct for the multiplicity of runs and architectures tested, so the reported numbers are optimistically biased. Because several margins in Table 4 are small (Spoken Digit: VQCCM 97.1 vs LEAF 96.7; Acoustic Scenes: 99.1 vs 99.1), selection bias could plausibly account for the claimed advantages. Please report the full distribution of results (e.g., mean and standard deviation over all runs used for selection), perform model selection on a held-out validation split that is disjoint from the test evaluation, or provide an explicit correction for selection.","section":"§5.1.1, §5.4.1, Table 4"},{"comment":"The abstract's statement that VQCCM 'achieves state-of-the-art (SOTA) performance across five different datasets' is contradicted by Table 4 itself: on Acoustic Scenes, TD-fbanks reaches 99.5 ± 0.4, which is higher than VQCCM's 99.1 ± 0.3, and VQCCM merely ties LEAF at 99.1. The sentence in §5.3.2 that 'VQCCM has outperformed LEAF for all tasks' is also false for Acoustic Scenes, where the means are equal (99.1 vs 99.1). The claims should be revised to specify the datasets on which VQCCM actually achieves a new best result, and the text in §5.3.2 should be made consistent with the numbers in Table 4.","section":"§5.3.2, Table 4, Abstract"},{"comment":"The architecture (filter sizes 100, 50, 12, 6, 3 and pooling sizes 10, 8, 4, 4) was selected on the S09 dataset for CosCovNN using the max-of-five protocol, and the same architecture is then used for the CNN baseline without any tuning of the CNN's filter sizes or pooling. The parameter-count reduction is a simple arithmetic fact, but the accuracy comparison between CosCovNN and CNN is confounded: the architecture search may have favored the cosine filters, and the CNN may not be at its best configuration. Please either run an equivalent architecture search for the CNN under the same protocol or compare both models on a fixed, pre-specified architecture and report the full distribution of results over runs.","section":"§5.1.1, §5.2.2, Table 4"}],"minor_comments":[{"comment":"The text from 'The field of audio classification has been fundamentally transformed...' through '...reduces the model's complexity by approximately 77' appears twice verbatim (once on pages 3–4 and once on pages 5–6). This duplication should be removed.","section":"Section 2"},{"comment":"Equation (7) writes 'Fi = F'i = Ek', which is confusing; it should define F'i = Ek with k = argmin_j ||Fi − Ej||2, and then state that the gradient is copied from F'i to Fi during backpropagation.","section":"Eq. (7)"},{"comment":"Reference [41] is cited for both CGCNN (Noé et al.) and CLDNN (Sainath et al.), which are different works, and the citation for TD-filterbanks in Table 1 is shown as an empty pair of brackets. The reference list and in-text citations should be corrected.","section":"References"},{"comment":"There are several typos and formatting issues: 'fare' should be 'fair' in §5.2.1; 'consisted of' should be 'consists of' in §4.1; 'comprises of' should be 'comprises' in §4.5; and the claim in §5.2.2 that cosine filters 'are less impacted by noise' is not supported by any experiment in the paper.","section":"General presentation"},{"comment":"The paper does not report whether code is available, and training hyperparameters such as learning rate, number of epochs, batch size, and optimizer are not specified, which makes the results difficult to reproduce or compare against.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The core idea is worth publishing if the empirical claims can be substantiated. The current version overstates the results via the SOTA claim and the max-of-five selection protocol. I encourage the editor to require a revision that (1) re-analyzes the data with proper validation and reports mean/median over runs, and (2) softens or corrects the SOTA claim for Acoustic Scenes. The duplicated related-work paragraphs also suggest the manuscript needs careful editorial cleanup before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the basic claim is simple and plausible. Replacing every convolution filter in a raw-audio CNN with a two-parameter cosine filter g[n]=θ1 cos(θ2n) is a clean, easy-to-check idea, and the parameter count math works (91,200 vs 408,192, a 77.66% reduction). CosCovNN beats the matched CNN on all five datasets in Table 4, and the ablation shows the memory module pulls its weight. If those numbers are real, this is a useful efficiency result for edge audio.\n\nThe main soft spot is the experimental protocol. Filter sizes, pooling sizes, memory size, and VQ codebook size are all chosen on S09 and IEMOCAP by taking the maximum accuracy over five training runs, and then the same datasets appear in the final comparison table. That is a selection-on-the-test-set problem. The margins in Table 4 are small enough (97.1 vs 96.7, 99.1 vs 99.1) that noise from the best-of-five pick could easily decide the outcome. The paper never reports what the typical run looks like. This is fixable with a held-out validation split or by reporting mean±std over seeds without model selection.\n\nSecond issue: the SOTA claim in the abstract is contradicted by the paper's own Table 4. TD-fbanks scores 99.5 on Acoustic Scenes, VQCCM scores 99.1, and the text later admits 'neither LEAF nor VQCCM could exceed TD-Fbanks performance in acoustic scene classification.' That inconsistency needs to be fixed. Third, the related work section contains two nearly identical duplicated paragraphs, and there are citation/formatting problems (e.g., one reference number serving both Noé and Sainath). These don't sink the idea, but they signal a manuscript that wasn't carefully assembled.\n\nOn the positive side, the work is a legitimate extension of parametric-filter frontends (SincNet, Gabor, LEAF) to every layer, and it is honest in places: the authors acknowledge CosCovNN doesn't beat LEAF on its own, and that the VQ layer alone doesn't help consistently. That gives me some confidence the authors are not deliberately cooking the results.\n\nWho this is for: anyone working on efficient raw-waveform classifiers or learnable filters. The paper deserves a serious referee, but the referee should insist on a clean evaluation protocol and a careful rewrite. I would not cite it in its current form, but I'd read a revised version.","headline":"A simple two-parameter cosine filter for every conv layer is a plausible efficiency idea, but best-of-five selection on the evaluation datasets and a self-contradicted SOTA claim make the reported gains unverified.","tokens_in":18357,"tokens_out":2889,"would_cite":false,"duration_ms":25942,"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":"Replacing CNN kernels with two-parameter cosine filters is claimed to make raw-audio classifiers both smaller and more accurate, with a memory-augmented version reaching state-of-the-art results.","keywords":["raw audio classification","cosine convolutional neural network","parametric filter design","learnable audio frontend","vector quantization","memory-augmented neural networks","speech emotion recognition","acoustic scene classification"],"falsifier":"Re-run the five-dataset comparisons with the same architectures but report the mean or median over five independent seeds instead of the maximum; if CosCovNN no longer beats the equivalent CNN on most tasks, or VQCCM no longer exceeds LEAF on typical runs, the central efficiency-accuracy claim fails. A complementary check is to move the S09 architecture search to a held-out validation split and see whether the chosen filter and pooling sizes still transfer to the other four datasets.","tokens_in":17238,"feed_emoji":"🎧","tokens_out":9886,"duration_ms":84135,"temperature":0.7,"pith_summary":"This paper sets out to show that a convolutional network for raw audio does not need thousands of learned filter weights: each filter can be generated by a cosine function with just two learnable parameters, an amplitude and a frequency. The resulting model, CosCovNN, is reported to be more accurate than an equivalently structured CNN on all five audio classification datasets tested, while using 77.66% fewer parameters. The paper further claims that adding a vector-quantisation layer and a learned memory module, producing VQCCM, lifts accuracy to state-of-the-art levels on those same datasets. If these results hold, they imply that building periodicity into the filter, rather than letting the network discover it from scratch, makes raw-waveform audio classifiers both smaller and more reliable.","feed_headline":"Cosine filters beat CNNs on raw audio with 77% fewer parameters","feed_subtitle":"Each filter is just a learnable cosine; memory plus quantisation lifts it to state of the art.","key_machinery":"The load-bearing object is the cosine filter $g[n]=\\theta_1\\cos(\\theta_2 n)$, a length-$L$ kernel generated by evaluating a periodic function with two trainable scalars: amplitude $\\theta_1$ and angular frequency $\\theta_2$. A standard 1D convolution would learn all $L$ tap values; CosCovNN instead fills the taps from this formula, so a layer's parameter cost is 2 per filter rather than $L$ per filter. The bounded periodic values are preserved through the network with tanh activations, and the classifier replaces a large fully connected layer with $Z$ cosine-convolutional layers followed by global average pooling. In VQCCM, the supporting mechanisms are a vector-quantisation codebook, which replaces each feature vector with its nearest codebook entry and copies gradients around the non-differentiable argmin, and a memory module whose learned vector is read, gated, multiplied into each layer's features, and written back additively.","core_discovery":"The central claim is that the standard learned kernel of a 1D CNN can be replaced, for audio, by a cosine filter $g[n]=\\theta_1\\cos(\\theta_2 n)$ without sacrificing accuracy—on the contrary, CosCovNN is reported to beat the equivalent CNN on all five tasks. Because every filter of length $L$ is specified by only two scalars, the convolutional parameter count falls from $L$ per filter to 2 per filter, which the paper computes as a 77.66% reduction for its five-layer architecture. The augmented model VQCCM inserts a vector-quantisation codebook after the first cosine layer and a memory reader-writer at every layer; the paper reports that VQCCM surpasses the LEAF learnable-frontend baseline on all five datasets, ties it on acoustic scenes at 99.1% with a smaller standard deviation, and achieves state-of-the-art accuracy in several categories.","pith_inferences":["A sharper comparison than the paper's would pit CosCovNN against a CNN with matched total parameter count rather than the same architecture; that would isolate whether the cosine form itself, rather than smaller model size, drives the accuracy gain.","Because a cosine filter is a continuous function, it can be evaluated at any kernel length for the same two parameters, which suggests a route to very long filters for low-frequency structure without extra training cost.","The two-parameter form is effectively a learned Fourier feature; initialising $\\theta_2$ from DCT frequencies or letting each channel carry its own frequency could make the layer behave like a trainable DCT basis for a wider range of signals.","The memory size was tuned on one-second utterances, but acoustic scenes are ten-second clips, so a direct test is whether a fixed memory dimension still suffices when input length grows tenfold."],"forward_implications":["If the cosine parametrisation is as effective as reported, any raw-waveform CNN can cut its convolutional parameter count by roughly the ratio of filter length to two, with no accuracy loss.","Because the architecture search was run once on the S09 spoken-digit dataset and then transferred to speech emotion, speaker identification, acoustic scenes, and music, the paper's recipe is claimed to be dataset-agnostic within audio.","The ablation study suggests that the memory module helps consistently, while vector quantisation alone does not, but the two together give the largest gains—so they appear to solve complementary problems.","A 77.66% parameter reduction makes CosCovNN and VQCCM natural candidates for memory-constrained and on-device audio classification pipelines."],"supporting_citations":[{"why":"Supplies the SincNet learnable band-pass filter baseline against which CosCovNN is compared on raw waveforms.","marker":"[9]"},{"why":"Supplies the LEAF learnable-frontend baseline that VQCCM claims to surpass on all five datasets.","marker":"[10]"},{"why":"Supplies the time-domain filterbank baseline and the experimental setup whose accuracy numbers are used for comparison.","marker":"[42]"},{"why":"Supplies the vector-quantisation loss and gradient-copy trick used to train the VQCCM quantisation layer.","marker":"[50]"},{"why":"Supplies global average pooling, which the classification layer uses to avoid a large fully connected layer.","marker":"[52]"},{"why":"Supplies the Speech Commands corpus from which the S09 spoken-digit dataset used for architecture search is derived.","marker":"[53]"},{"why":"Supplies the IEMOCAP speech emotion dataset used in the main and ablation comparisons.","marker":"[54]"},{"why":"Supplies the VoxCeleb speaker identification dataset used in the main and ablation comparisons.","marker":"[55]"},{"why":"Supplies the TUT Urban Acoustic Scenes dataset used in the main and ablation comparisons.","marker":"[56]"},{"why":"Supplies the NSynth musical instrument dataset used in the main and ablation comparisons.","marker":"[57]"}],"fun_headline_variants":["Cosine filters cut CNN parameters by 77% for raw audio","Audio CNN: two-parameter cosine filters beat full kernels","CosCovNN: 77% fewer parameters, state-of-the-art audio","VQCCM: cosine CNN with memory tops five audio datasets","Leaner CNNs: cosine filters for raw audio classification"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The protocol records the maximum accuracy over five training runs and uses those maxima to choose filter and pooling sizes on the S09 dataset and to compare models, so if the best of five runs is a lucky outlier rather than the typical outcome, the reported margins over the CNN baseline and over LEAF could shrink.","fun_headline_variants_meta":{"raw":{"variants":["Cosine filters cut CNN parameters by 77% for raw audio","Audio CNN: two-parameter cosine filters beat full kernels","CosCovNN: 77% fewer parameters, state-of-the-art audio","VQCCM: cosine CNN with memory tops five audio datasets","Leaner CNNs: cosine filters for raw audio classification"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000461,"raw_usage":{"total_tokens":2286,"prompt_tokens":905,"completion_tokens":1381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":1292}},"tokens_in":521,"tokens_out":1381,"duration_ms":10188,"temperature":1.0,"reasoning_tokens":1292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:30:30.401227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the five-dataset comparisons with the same architectures but report the mean or median over five independent seeds instead of the maximum; if CosCovNN no longer beats the equivalent CNN on most tasks, or VQCCM no longer exceeds LEAF on typical runs, the central efficiency-accuracy claim fails. A complementary check is to move the S09 architecture search to a held-out validation split and see whether the chosen filter and pooling sizes still transfer to the other four datasets.","supporting_citations":[{"cited_title":"Speaker recognition from raw waveform with sincnet,","cited_arxiv_id":null,"evidence_quote":"Supplies the SincNet learnable band-pass filter baseline against which CosCovNN is compared on raw waveforms."},{"cited_title":"LEAF: A Learnable Frontend for Audio Classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the LEAF learnable-frontend baseline that VQCCM claims to surpass on all five datasets."},{"cited_title":"Learning filterbanks from raw speech for phone recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the time-domain filterbank baseline and the experimental setup whose accuracy numbers are used for comparison."},{"cited_title":"Neural discrete representa- tion learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the vector-quantisation loss and gradient-copy trick used to train the VQCCM quantisation layer."},{"cited_title":"IEMOCAP: Interactive emotional dyadic motion capture database,","cited_arxiv_id":null,"evidence_quote":"Supplies the IEMOCAP speech emotion dataset used in the main and ablation comparisons."},{"cited_title":"A multi-device dataset for urban acoustic scene classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the TUT Urban Acoustic Scenes dataset used in the main and ablation comparisons."},{"cited_title":"Neural audio synthesis of musical notes with wavenet au- toencoders,","cited_arxiv_id":null,"evidence_quote":"Supplies the NSynth musical instrument dataset used in the main and ablation comparisons."}],"review_version":1}