{"id":"66d654b6-4d05-4e65-9550-98942144f833","arxiv_id":"2509.10031","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A generic 2D convolutional front-end trained from raw audio matches existing learnable front-ends on LibriSpeech while using far fewer parameters.","lead":"This paper shows that a speech recognizer can learn its own acoustic features using only a layered stack of 2D convolutions, with no hand-designed filterbanks. The result is a smaller, cheaper front-end that matches existing learned feature extractors on LibriSpeech, though still slightly behind log Mel.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parity claim rests on under-tuned baselines; re-running SCF/wav2vec with full audio perturbation may eliminate it.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing concern: the comparison baselines in Table 1 are under-tuned. The paper's central claim of parity with existing learnable feature extractors is supported only by these weakened baselines. The paper itself acknowledges this in Section 4 and Section 5, but the abstract and conclusions do not carry the same caveat, making the claim overbroad. I considered other potential concerns, such as single-run WERs and configuration selection on dev-other, but these are secondary: even if repeated runs and held-out selection were performed, the baseline tuning issue would still undermine the parity claim. The parameter-efficiency result is separate and largely unaffected by this concern, but the headline 'matches existing supervised learnable feature extractors' is conditional on the baseline setup. The proposed concrete test—re-running the baselines with the perturbation recipe from [23]—would settle whether the parity persists under properly tuned baselines. Until then, the conditional verdict is appropriate; no change from the reader's disposition is needed.","tokens_in":9625,"tokens_out":4649,"duration_ms":52935,"concrete_test":"Re-run the SCF and wav2vec baselines from Table 1 using the audio perturbation recipe of [23] (e.g., tempo perturbation with factors sampled from a continuous distribution, plus the same SpecAugment settings), while keeping the 2D front-end and all other hyperparameters fixed. If either baseline improves by more than ~0.2% absolute WER on dev-other relative to Table 1, the parity claim is not robust; if they do not improve, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a generic 2D convolutional front-end 'performs on par with SCF' and 'matches the performance of existing supervised learnable feature extractors' (Abstract, Sec. 4). The evidence is Table 1, where the 2D front-end (2.3M) reaches 2.5/5.5 dev and 2.9/5.9 test, versus SCF at 2.6/5.7 and 2.9/6.0. However, Section 4 explicitly states that, unlike prior work [23], 'we do not tune the audio perturbation and only applied speed perturbation with rather limited perturbation factors,' and attributes the degraded SCF/wav2vec results to this. The differences between the 2D front-end and SCF are at most 0.2% absolute. If SCF or wav2vec were trained with the stronger tempo perturbation used in [23], published results (e.g., [11,20]) show they can match or exceed log Mel, while the 2D front-end remains below log Mel by ~0.3% on dev-other. The abstract's phrase 'existing supervised learnable feature extractors' invites comparison to published SCF/wav2vec numbers, not the paper's own weakened baselines. Thus the parity result is conditional on an under-tuned comparison, and the claim that the generic architecture is sufficient to match existing extractors is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified, randomly initialized 2D convolutional front-end for ASR that replaces both the handcrafted log-Mel/SCF feature extraction and the subsequent VGG-style subsampling block with a single stack of 2D convolutions. On LibriSpeech with a CTC/Conformer acoustic model, the proposed 2D front-end achieves test WERs of 2.9/5.9 (clean/other) versus 2.9/6.0 for the paper's SCF baseline, while using far fewer parameters before the Conformer (2.3M or 0.3M versus 12.4M for SCF and 5.0M for wav2vec). The paper includes ablations on the first-layer type (STFT vs. filterbank, random vs. Gammatone initialization), the number of 2D layers, subsampling factors, kernel size, and channel count, plus an analysis of learned filter frequency responses. The authors conclude that a generic, from-scratch front-end can match existing supervised learnable feature extractors while being parameter-efficient.","tokens_in":9981,"tokens_out":9999,"duration_ms":108428,"significance":"If the parity claim holds under properly tuned training conditions, the contribution is significant: it challenges the necessity of handcrafted inductive bias in neural front-ends and provides a parameter-efficient alternative for resource-constrained ASR. The paper is clearly written, contains detailed and fairly systematic ablations (Tables 2, Figures 1 and 3), and releases code. The learned-filter analysis is a useful qualitative check. However, the significance is conditional on the central comparison being fair, and the paper itself acknowledges that the baselines are under-tuned relative to prior work. The main scientific value is the architecture and the ablation evidence; the headline claim of matching existing extractors is not yet fully established.","major_comments":[{"comment":"The central parity claim ('matches the performance of existing supervised learnable feature extractors') is not yet supported by the evidence. The paper states in §4 that audio perturbation was not tuned and only speed perturbation with factors {0.9, 1.0, 1.1} was used, which the authors themselves say may explain why the SCF and wav2vec baselines underperform their published results in [11, 20]. The 2D front-end's best WERs are within 0.1–0.2% absolute of this under-tuned SCF baseline. If SCF/wav2vec were retrained with the stronger perturbation used in prior work, they could match or exceed log Mel, while the 2D front-end remains below log Mel in this paper. The abstract should either restrict the claim to 'under the same training recipe' or the authors should retrain the baselines with matched, tuned augmentation.","section":"Abstract; §4, Table 1"},{"comment":"The parity conclusion rests on WER differences of 0.1–0.2% absolute (e.g., 2.9 vs. 2.9 on test-clean and 5.9 vs. 6.0 on test-other). No multiple seeds or confidence intervals are reported. In CTC ASR, such small differences are often within run-to-run variation. Please report multiple runs with mean and standard deviation, or clearly state that the differences are not statistically distinguished. This is load-bearing because the paper's main claim is that the 2D front-end is 'on par with SCF.'","section":"Table 1; §4"},{"comment":"The conclusion that 'the remaining difference to log Mel features is not due to its missing structure' is an overreach. The evidence is only that the proposed 2D front-end is competitive with under-tuned SCF/wav2vec baselines under the same weak augmentation. The paper itself lists weak audio perturbation as a possible cause of the gap to log Mel in §5. A controlled comparison with a properly tuned log-Mel baseline is needed to attribute the gap to structure rather than to insufficient regularization or tuning effort. The current wording overstates what the experiments show.","section":"§6; §5"},{"comment":"The parameter-efficiency comparison is not fully transparent. The reported count 'parameters before the Conformer' includes the linear projection to the Conformer dimension and, for log Mel and SCF, the VGG-style subsampling block, while the proposed 2D front-end subsumes the subsampling. Thus the large parameter saving (12.4M vs. 2.3M/0.3M) partly reflects the lower-dimensional output representation and the architectural unification, not purely the feature extractor itself. Please provide a breakdown of feature-extractor, subsampling/VGG, and projection parameters, and clarify in the text that the comparison is for the entire front-end including projection.","section":"Table 1; §2.2"}],"minor_comments":[{"comment":"The STFT row lacks a parameter count, and the SpecAugment column entries ('Features' vs. 'STFT'?) are ambiguous. Please align the columns and clarify what is being reported.","section":"Table 1"},{"comment":"The table layout is garbled; it is unclear which rows correspond to trainable vs. fixed and Gammatone vs. random initialization. Please reformat the table.","section":"Table 2"},{"comment":"The axis labels and cell values are hard to read in the text version. A heatmap with explicit axes and a legend would be clearer.","section":"Figure 1"},{"comment":"The text says that dropping to kernel size 16 'introduces a severe impairment', but the figure caption says the model with kernel size 16 did not converge. Please reconcile; if it did not converge, the WER should not be presented as a valid comparison point.","section":"Figure 3"},{"comment":"The characterization of the front-end as 'generic' and 'from scratch' should be qualified: the first-layer kernel size, stride, and number of channels are tuned (Section 2.2, Figure 3), and the architecture still uses a filterbank-like or STFT-like first layer to create the feature dimension. The claim should say 'randomly initialized' rather than implying no handcrafted influence at all.","section":"Abstract; §2.2"},{"comment":"The claim that the learned filters show 'clear groups of filters with adjacent ascending or descending center frequencies' is qualitative and based on visual inspection. Consider providing a quantitative measure of ordering or a more systematic analysis.","section":"§4.1"},{"comment":"The code link is a bare repository path. Please include a commit hash or version and a brief description of how to run the experiments.","section":"Reproducibility"},{"comment":"Reference [23] is marked 'to appear'. Please update it with the final publication details if available.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental but solid experimental study. The main risk is overclaiming parity under under-tuned baselines. If the authors retune the baselines or soften the claims and add seed variability, the contribution would be acceptable for publication. No ethical concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take.\n\nThe paper is a careful, well-scoped empirical study. The actual new thing is the unified 2D convolutional front-end: one stack of 3x3 2D convs applied directly to the raw waveform, with SpecAugment moved into the STFT domain instead of between a 1D feature extractor and a VGG-style block. That's a genuine simplification over SCF and wav2vec-style front-ends, which splice 1D waveform filters onto a separate 2D subsampling block. The ablations are systematic—first-layer type, learnable vs frozen, random vs Gammatone init, subsampling placement, kernel size, channel count—and the learned-filter analysis (bandpass shapes, adjacent frequency ordering induced by 3x3 kernels) is a nice observation. Parameter efficiency is real: 0.3–2.3M parameters before the Conformer vs 12.4M for SCF and 5.0M for wav2vec, with roughly matching WERs in their setup. Code is public, reproducibility on a single GPU is a plus.\n\nThe soft spot is exactly what the stress-test note flags. Section 4 says audio perturbation was not tuned (only speed perturbation with factors {0.9,1.0,1.1}), and that this likely explains why SCF and wav2vec underperform their published numbers. So the abstract's phrase 'matches the performance of existing supervised learnable feature extractors' is too strong. What is shown is parity with the authors' own weakly-regularized baselines. Whether that parity holds when SCF and wav2vec are trained with the stronger tempo perturbation from their earlier work is unknown. The reader's conditional verdict is right.\n\nAlso, WERs are single runs with no significance tests. Differences of 0.1–0.2% absolute are within noise, and the 2D configuration was selected on dev-other, so test numbers are mildly cherry-picked. Minor, but worth noting.\n\nCredit where due: the paper is honest about all of this in Section 5, and there's no circularity—the comparison is empirical, baselines are external, and the 2D front-end is trained from scratch.\n\nWho it's for: people working on learnable front-ends, raw-waveform ASR, or low-resource scenarios. Not a breakthrough, but a useful data point and a clean ablation study.\n\nRecommendation: send it to peer review. A serious referee should ask for re-running SCF and wav2vec with the stronger perturbation and ideally multiple seeds. With that, the central claim would be solid. As it stands, the architecture is plausible and the paper is appropriately cautious about its limitations.","headline":"A genuine architectural simplification for learnable front-ends, but the parity claim is undersold by weak baselines.","tokens_in":10515,"tokens_out":3175,"would_cite":true,"duration_ms":32538,"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":"A single 2D convolutional front-end, trained from scratch, matches handcrafted-inspired feature extractors for ASR.","keywords":["automatic speech recognition","learnable front-end","2D convolution","feature extraction","raw waveform modeling","parameter efficiency","CTC","Conformer"],"falsifier":"Re-run Table 1 with SCF and wav2vec baselines trained using continuous tempo perturbation with stronger factors (as in prior work, reference [23]), keeping everything else fixed. If SCF drops below 2.3/5.0 WER on dev clean/other while the 2D front-end stays at 2.5/5.5, the paper's central parity claim fails.","tokens_in":9503,"feed_emoji":"🎙️","tokens_out":2972,"duration_ms":30192,"temperature":0.7,"pith_summary":"The paper tries to show that a generic, unified 2D convolutional front-end, trained from scratch on labeled speech, can replace the handcrafted feature extraction pipeline and the separate VGG-style subsampling block in an ASR model. It claims this is possible without inheriting classical design constraints like Gammatone-inspired filter shapes or fixed log/root nonlinearities. On LibriSpeech with a CTC-Conformer model, the proposed front-end matches the word error rate of SCF and wav2vec-style extractors while using far fewer parameters (0.3M–2.3M vs 5.0M–12.4M). The authors conclude that most of the inductive bias in audio front-ends is unnecessary.","feed_headline":"One 2D conv front-end matches handcrafted speech features","feed_subtitle":"Trained from scratch on LibriSpeech, it matches SCF and wav2vec extractors with 5–40x fewer parameters.","key_machinery":"The core object is a unified 2D convolutional front-end: a first layer that turns the 1D waveform into a time-frequency representation (either STFT magnitude or a learnable filterbank), followed by six 3x3 2D convolutional layers with stride 2 in time, each doubling subsampling until the 40 ms frame rate. This single stack absorbs both the classical feature extraction and the VGG-style subsampling block, and it is trained jointly with the acoustic model from scratch.","core_discovery":"On its own terms, the paper's central claim is that a single stack of 2D 3x3 convolutions, applied directly to the waveform through a learnable first-layer filterbank, can serve as the entire feature extraction front-end for ASR and perform on par with established learnable front-ends. The generic design removes the arbitrary split between 1D waveform feature extraction and 2D VGG-style subsampling, and moves SpecAugment to the STFT domain before the front-end. The result: WERs of 2.5/5.5 on dev clean/other and 2.9/5.9 on test clean/other, statistically on par with SCF (2.6/5.7, 2.9/6.0) and better than the wav2vec extractor in this setup, at a fraction of the parameter count. Analyses of le","pith_inferences":["If the parity holds, classical Gammatone/log-Mel structure is not a necessary inductive bias for good ASR features; the model can learn equivalent structure directly from the task loss. A testable extension would be applying this generic front-end to other audio tasks and comparing against task-specific front-ends.","The observed filter-ordering (adjacent ascending/descending center frequencies) suggests the 2D convolutions impose a locality prior on the feature dimension; initializing the filterbank in a frequency-sorted order might accelerate convergence or improve the small-channel configuration.","The wav2vec-style 1D extractor may be superseded by the simpler 2D stack in resource-constrained settings; a direct comparison on a larger corpus or with a Transformer-encoder acoustic model would test this more broadly."],"forward_implications":["ASR front-ends can be built from a single generic 2D convolutional stack, removing the architectural patchwork of 1D front-end plus VGG block.","Parameter count before the Conformer drops from 12.4M (SCF) or 5.0M (wav2vec) to as little as 0.3M, making fully learnable front-ends feasible on a single 24GB GPU.","Moving SpecAugment to the STFT domain makes its hyperparameters independent of the feature extractor's output dimension, so the same augmentation works across different front-ends.","The learned filterbank still develops Gammatone-like bandpass filters, showing a generic design can rediscover classical auditory filters without explicit prior structure.","The parameter-efficient configuration (0.3M) loses only about 0.3-0.4 WER absolute versus the larger one, offering a clear compute-accuracy trade-off."],"fun_headline_variants":["Unified 2D conv front-end matches SCF and wav2vec","Single 2D conv stack matches learnable speech extractors","Parameter-efficient 2D conv front-end rivals wav2vec","Generic 2D conv features replace split 1D+2D pipelines","2D conv front-end from scratch matches pre-trained extractors"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The parity claim rests on the SCF and wav2vec baselines being limited by the authors' weak audio perturbation; with the stronger perturbation settings used in prior work, those baselines improve, and the 2D front-end's parity could disappear.","fun_headline_variants_meta":{"raw":{"variants":["Unified 2D conv front-end matches SCF and wav2vec","Single 2D conv stack matches learnable speech extractors","Parameter-efficient 2D conv front-end rivals wav2vec","Generic 2D conv features replace split 1D+2D pipelines","2D conv front-end from scratch matches pre-trained extractors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000422,"raw_usage":{"total_tokens":1996,"prompt_tokens":730,"completion_tokens":1266,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":1172}},"tokens_in":474,"tokens_out":1266,"duration_ms":11603,"temperature":1.0,"reasoning_tokens":1172,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:15:32.452398+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 1 with SCF and wav2vec baselines trained using continuous tempo perturbation with stronger factors (as in prior work, reference [23]), keeping everything else fixed. If SCF drops below 2.3/5.0 WER on dev clean/other while the 2D front-end stays at 2.5/5.5, the paper's central parity claim fails.","supporting_citations":[],"review_version":1}