{"id":"e4278e37-1c37-4ead-98be-c2456ad707c4","arxiv_id":"2502.00310","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A learnable wavelet network with asymmetric hard thresholding reports top accuracies on IEMOCAP and EMO-DB, though its speaker-independent evaluation protocol is questionable.","lead":"SigWavNet is a deep learning model that learns its own wavelet filters and denoising thresholds directly from raw audio, reporting 84.8% accuracy on IEMOCAP and 90.1% on EMO-DB for speech emotion recognition. It claims to beat prior methods without manual feature extraction, which matters for emotion-aware interfaces and mental-health screening, but the speaker-independent evaluation is undermined by a random data split.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'speaker-independent' evaluation is not speaker-disjoint: Section IV-B describes stratified random utterance-level splits, and on 10-speaker IEMOCAP/EMO-DB this allows speaker identity leakage, so the SOTA claims are not established.","rationale":"The reader's weakest assumption correctly identifies the core problem: the described evaluation is not speaker-independent because it uses stratified random utterance-level splits on datasets with only 10 speakers. This is not a matter of differing community convention; it is an internal inconsistency between the paper's explicit 'speaker-independent (SI)' claim in Section IV-B and the split procedure it actually describes. The consequence is a plausible large upward bias in the reported numbers, since speaker identity is a strong confound in emotion recognition. The central claim of outperforming state-of-the-art methods is therefore not supported by the evidence as presented. I do not see another concern with greater impact: while the comparison tables use published numbers under heterogeneous protocols, that weakness is secondary to the leakage issue, and the architecture itself is not the reason to reject. If the speaker-disjoint rerun reproduces the results, the verdict would need to be revisited; but based on the manuscript text, the reader's REJECT verdict is the appropriate call. The test is concrete, feasible, and directly settles the concern.","tokens_in":25060,"tokens_out":2045,"duration_ms":23564,"concrete_test":"Rerun the full SigWavNet pipeline on IEMOCAP and EMO-DB using leave-one-speaker-out cross-validation (train on 9 speakers, test on the held-out speaker), keeping the same architecture, hyperparameters, and loss. If the accuracy and F1 remain near 84.8%/85.1% and 90.1%/90.3%, the speaker-independence claim survives and the concern is refuted. If the numbers drop materially, the reported results are inflated by speaker leakage, and the paper's central claim should be revised. As a complementary check, inspect the provided GitHub repository's data-split code to verify whether speaker IDs are respected; the paper text currently indicates they are not.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SigWavNet achieves state-of-the-art speaker-independent SER accuracy (84.8% on IEMOCAP, 90.1% on EMO-DB, Tables II and IV). The paper labels these experiments 'speaker-independent (SI)' in Section IV-B, but the protocol described there is a stratified random split: 'the distribution within the cross-validation folds, was conducted using stratified random sampling' with strata based on emotion classes. No step separates utterances by speaker. Both datasets have only 10 speakers (IEMOCAP: five sessions, two speakers each; EMO-DB: five male and five female actors). With a 90/10 train/test partition and 10-fold cross-validation, the same speaker's utterances almost certainly appear in both training and test folds. A model can then exploit speaker-specific vocal traits to recognize emotion, inflating accuracy relative to a genuine speaker-independent protocol. Because the headline claim is superiority over prior methods specifically in the speaker-independent setting, this evaluation gap is load-bearing: it does not establish the claimed generalization to unseen speakers. The architecture and ablation study are coherent, but they evaluate a different, easier question. The comparison tables also mix protocols from published papers, so even the relative ranking is not controlled. The single most decisive weakness is the mismatch between the stated SI claim and the actual split described.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SigWavNet, an end-to-end multiresolution deep learning architecture for speech emotion recognition (SER). The model operates on raw waveforms, using a learnable fast discrete wavelet transform (FDWT) with learnable asymmetric hard thresholding (LAHT), followed by a 1D dilated CNN with spatial attention and a Bi-GRU with temporal attention, then channel weighting and global average pooling before classification. The authors report state-of-the-art speaker-independent accuracy of 84.8% on IEMOCAP and 90.1% on EMO-DB, with F1-scores of 85.1% and 90.3%, respectively. The paper includes a per-class performance analysis, a comparison with prior methods, and an ablation study examining the contributions of the learnable wavelet kernels, LAHT, and the attention components. Source code is provided via a GitHub repository.","tokens_in":25317,"tokens_out":3399,"duration_ms":34474,"significance":"The architectural ideas in this paper are interesting and potentially useful: a learnable wavelet filter bank that adapts during training, a differentiable asymmetric hard-thresholding activation, and the combination of multiresolution decomposition with attention-based sequence modeling are well-motivated. The ablation study is systematic and informative, showing the incremental contributions of LAHT, kernel learning strategies, and the Bi-GRU/temporal-attention module. The open-source release of the code is a strength. However, the central empirical claim—superiority over state-of-the-art methods in speaker-independent SER—is not supported by the evaluation protocol described in Section IV-B. Because the data splits are stratified random utterance-level splits rather than speaker-disjoint splits, and both datasets have only 10 speakers, the reported accuracies likely benefit from speaker identity leakage. If the experiments were redone with a proper speaker-independent protocol, the method could be a valuable contribution, but as presented the headline result is unsubstantiated.","major_comments":[{"comment":"The evaluation protocol is not speaker-independent despite being labeled as such. The text states that data are partitioned using \"stratified random sampling\"[55] with strata based on emotion classes, and no step ensures that utterances from the same speaker are confined to either training or test sets. IEMOCAP and EMO-DB each contain only 10 speakers; with a 90/10 stratified random split, the same speaker's utterances will almost certainly appear in both training and test folds. Under these conditions, the model can exploit speaker-specific vocal characteristics to predict emotion, inflating accuracy relative to a true speaker-independent evaluation. Consequently, the claimed speaker-independent accuracies of 84.8% (IEMOCAP) and 90.1% (EMO-DB) in Tables II and IV do not establish generalization to unseen speakers. The authors should re-run the experiments with speaker-disjoint splits, for example leave-one-speaker-out or grouped k-fold by speaker, and report results under that protocol.","section":"Section IV-B"},{"comment":"The comparison with state-of-the-art methods is not controlled across protocols. The manuscript does not state whether each cited method in Tables II and IV used a speaker-independent, speaker-dependent, or random split, nor does it specify the exact folds or cross-validation procedures for those methods. On small datasets like IEMOCAP and EMO-DB, protocol differences can substantially change reported performance. For example, if some compared methods used a more difficult speaker-disjoint protocol while SigWavNet used a random split, the relative ranking in these tables would be misleading. The authors should compare all methods under an identical speaker-disjoint protocol, or at minimum transparently report the evaluation protocol of each compared work and discuss the compatibility of the numbers.","section":"Tables II and IV"},{"comment":"The reported gains lack measures of variability. The text says each test is repeated 10 times with different random seeds, but Tables I-IV report only a single accuracy and F1-score with no standard deviation, confidence interval, or per-fold results. With only 10 speakers and random splits, the variance across seeds could be large, and the difference between 84.8% and the next-best IEMOCAP result of 80.8% may not be statistically significant. Reporting mean and standard deviation across the 10 repetitions, or per-fold results under a speaker-disjoint split, is needed to support the claim of outperforming prior methods.","section":"Section IV-B and Tables I-IV"}],"minor_comments":[{"comment":"The phrase \"dilated 1D dilated convolutional layers\" is redundant; it should be \"dilated 1D convolutional layers\".","section":"Section III-C"},{"comment":"The condition \"α · β < 0\" appears inconsistent with the statement that to emulate linear activation the biases can be set to zero \"while allowing α and β to be real numbers\"; if α and β are both positive, their product is positive, violating the condition. Please clarify the intended range of α and β.","section":"Equation (16)"},{"comment":"The sentence \"totaling 10039 utterances each\" is unclear; IEMOCAP has 10,039 utterances in total, not per speaker or per session. Please rephrase.","section":"Section IV-A.1"},{"comment":"The caption \"Daubechies Wavelets kernel size\" does not explain what is plotted (e.g., the scaling and wavelet function, or the filter coefficients). Clarify the figure and its relation to the initialization of Conv_h and Conv_g.","section":"Figure 7"},{"comment":"Reference [2] is a bare URL for a CDC page; it should be formatted consistently with the other references (author, title, year, and access date).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The evaluation-protocol problem is the central issue. The architecture and ablation are coherent, and the authors provide code, so the work could become publishable after a proper speaker-independent evaluation. I would recommend requiring the authors to re-run all experiments with speaker-disjoint folds and to redo the comparison tables under that protocol. If the authors are unwilling or unable to provide such results, the paper should not be accepted, because the headline claim would remain unsupported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper is simple: the model design deserves a look, but the headline claim does not survive contact with the evaluation section. They report 84.8% on IEMOCAP and 90.1% on EMO-DB and call it speaker-independent, but Section IV-B describes a stratified random split by emotion class, not a speaker-disjoint split. IEMOCAP and EMO-DB have ten speakers each. With random utterance-level folds, the same speakers are almost certainly in both training and testing, so the numbers likely reflect speaker identity leakage. The stress-test note is right, and the paper's own text confirms it.\n\nWhat is actually new: the learnable FDWT front-end with the CQF constraint, the LAHT activation with learnable asymmetric thresholds, and the full end-to-end stack are a coherent, plausible combination. Prior wavelet-learnable work exists, but the asymmetric thresholding is a genuine twist. Credit where due: the ablation study is informative, the Daubechies initialization is sensible, and they shared code and a GitHub link. That is real evidence of reproducibility intent.\n\nSoft spots, in proportion. The evaluation flaw is load-bearing, not cosmetic. The comparison tables also pull published numbers from other papers without a shared protocol, so even the ranking relative to prior work is uncontrolled. No error bars are reported, which matters less but is worth noting. A minor point: the paper claims \"eliminating the need for pre or post-processing\" but still standardizes sampling rate and mono conversion; that is acceptable, but the phrasing oversells.\n\nIf you strip out the SI claim, what remains is a plausible architecture with a suggestive ablation on a poorly controlled test. That is not enough to establish state-of-the-art. The fix is straightforward: rerun with leave-one-speaker-out or speaker-disjoint folds. If the margins hold, this becomes a solid contribution. As written, the main claim is unverified.\n\nWho gets value: anyone working on learnable wavelet front-ends or SER evaluation pitfalls. The paper is a good teaching example of why split strategy matters on small-speaker datasets. It deserves a serious referee, but the referee should treat the current results as provisional and require major revision.\n\nBottom line: send it to review, but expect the authors to redo the evaluation under a proper speaker-independent protocol before the numbers can be believed.","headline":"The architecture is interesting and the ablation is useful, but the speaker-independent claim is not supported by the described stratified random split on 10-speaker datasets.","tokens_in":25889,"tokens_out":1268,"would_cite":false,"duration_ms":15177,"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":"SigWavNet proposes a fully learnable multiresolution wavelet network for speech emotion recognition, reporting accuracy and F1 gains over prior methods on IEMOCAP and EMO-DB.","keywords":["speech emotion recognition","fast discrete wavelet transform","learnable wavelet filters","asymmetric hard thresholding","dilated CNN","Bi-GRU","attention mechanism","IEMOCAP"],"falsifier":"Run SigWavNet with leave-one-speaker-out cross-validation on IEMOCAP and EMO-DB, training on nine speakers and testing on the held-out tenth, and compare accuracy and F1 with the reported 84.8%/85.1% and 90.1%/90.3%; a noticeably lower score would indicate that the original numbers rely on speaker leakage.","tokens_in":24836,"feed_emoji":"🎙️","tokens_out":5452,"duration_ms":50407,"temperature":0.7,"pith_summary":"This paper tries to establish that a speech emotion recognition system can be built end-to-end from raw waveforms by replacing the fixed wavelet transform with a learnable fast discrete wavelet transform, and that this design beats previously published methods on two standard acted-speech datasets. The authors report speaker-independent accuracy of 84.8% and F1 of 85.1% on IEMOCAP, and 90.1% accuracy and 90.3% F1 on EMO-DB, without fixed-length segmentation or separate denoising. If true, the result suggests that wavelet-inspired filter banks can be trained as part of a neural network rather than hand-designed, and that such a front end helps emotion recognition by localizing cues in both time and frequency.","feed_headline":"Learnable wavelets beat prior speech-emotion models","feed_subtitle":"SigWavNet reads raw waveforms and claims 84.8% on IEMOCAP and 90.1% on EMO-DB without fixed-length segmentation.","key_machinery":"The load-bearing mechanism is the learnable fast discrete wavelet transform (FDWT) layer: a cascade of convolutional filter pairs, initialized as Daubechies-10 wavelet coefficients and refined by training, that decompose raw speech into approximation and detail coefficients at several levels. The conjugate quadrature filter identity $g[n]=(-1)^n h[-n]$ derives the high-pass kernel from the low-pass kernel, halving parameters while preserving orthogonality. A learnable asymmetric hard-thresholding (LAHT) activation applies two sigmoid-based gates with separate positive and negative biases to mimic wavelet denoising inside the network. Downstream, 1D dilated convolutions with spatial attention and a bidirectional GRU with temporal attention extract spatial and sequential patterns from each band, and a channel-weighting layer reweights the bands before global average pooling and log-softmax classification.","core_discovery":"The central claim is that making the discrete wavelet transform itself trainable removes the main obstacle to using wavelets in deep learning for speech emotion recognition. SigWavNet treats the low-pass and high-pass filters of each decomposition level as convolutional kernels initialized with Daubechies-10 coefficients, constrains the high-pass kernel to be the alternating flip of the low-pass kernel through the conjugate quadrature filter property, and applies a learnable asymmetric hard-thresholding activation to the detail coefficients at each level. The resulting multiresolution coefficients are fed to a dilated CNN with spatial attention and a Bi-GRU with temporal attention, then pooled and classified with log-softmax. The authors argue that this learns a data-driven filter bank resembling a Mel-scale emphasis on low frequencies while preserving high-frequency details, and they report that the full model outperforms all compared methods on both datasets.","pith_inferences":["The evaluation claim should be tested with leave-one-speaker-out splits. IEMOCAP and EMO-DB contain only ten speakers each, so stratified random utterance splits almost certainly mix the same speaker into training and test, and the reported 'speaker-independent' numbers may partly reflect speaker-identity cues; a speaker-disjoint evaluation would show how much of the margin survives.","A natural extension is a noise-robustness test: add babble, music, or channel noise to the raw waveforms and measure whether the learnable thresholding maintains its advantage over fixed hard thresholding.","The learnable FDWT idea is not limited to emotions; it could be transferred to other variable-length physiological and acoustic signals, such as ECG, EEG, or environmental audio, where learnable multiscale denoising is likely to behave similarly."],"forward_implications":["If the reported results hold under proper speaker separation, raw-waveform SER no longer needs fixed-length segmentation, hand-designed features, or a separate denoising stage.","The learnable FDWT front end offers a compact inductive bias: an $L$-level cascade uses only $2L$ filters, far fewer than typical 1D CNNs, which could make the architecture attractive for on-device emotion recognition.","The same wavelet-to-CNN design could be applied to other 1D signal classification tasks where multiscale denoising and time-frequency localization matter.","The reported margins over prior methods on IEMOCAP and EMO-DB would make SigWavNet the default baseline to beat in future SER comparisons, assuming the evaluation protocol is confirmed."],"supporting_citations":[{"why":"IEMOCAP dataset: provides the primary benchmark and the label protocol (happy/excited merged) used in all evaluations.","marker":"[51]"},{"why":"EMO-DB dataset: supplies the second benchmark and its seven emotion labels.","marker":"[54]"},{"why":"Mallat's wavelet tour: gives the FDWT/cascade algorithm and orthonormal wavelet basis theory the learnable FDWT layer is built on.","marker":"[47]"},{"why":"The conjugate quadrature filter concept that yields the g = alternating flip of h constraint and parameter reduction.","marker":"[48]"},{"why":"Daubechies wavelets used to initialize the learnable filters and motivate time-frequency localization.","marker":"[33]"},{"why":"Focal loss with L2 regularization: the training objective used to handle class imbalance and hard examples.","marker":"[58]"},{"why":"Stratified random sampling: the splitting procedure whose speaker-independence is the load-bearing evaluation assumption.","marker":"[55]"},{"why":"Network in network: motivation for global average pooling instead of fully connected layers.","marker":"[49]"}],"fun_headline_variants":["Learnable wavelets boost speech emotion recognition","SigWavNet trains wavelet filters for emotion detection","End-to-end SER with trainable wavelet transforms","Multiresolution wavelet network outdoes SER state-of-the-art","Trainable wavelet denoising lifts speech emotion accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that stratified random utterance-level splits give a speaker-independent evaluation, even though each dataset has only ten speakers, so the same speaker's voice almost certainly appears in both training and test folds.","fun_headline_variants_meta":{"raw":{"variants":["Learnable wavelets boost speech emotion recognition","SigWavNet trains wavelet filters for emotion detection","End-to-end SER with trainable wavelet transforms","Multiresolution wavelet network outdoes SER state-of-the-art","Trainable wavelet denoising lifts speech emotion accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1455,"prompt_tokens":1014,"completion_tokens":441,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":366}},"tokens_in":630,"tokens_out":441,"duration_ms":4280,"temperature":1.0,"reasoning_tokens":366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:28:10.848008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SigWavNet with leave-one-speaker-out cross-validation on IEMOCAP and EMO-DB, training on nine speakers and testing on the held-out tenth, and compare accuracy and F1 with the reported 84.8%/85.1% and 90.1%/90.3%; a noticeably lower score would indicate that the original numbers rely on speaker leakage.","supporting_citations":[{"cited_title":"Mallat, A Wavelet Tour of Signal Processing: The Sparse Way","cited_arxiv_id":null,"evidence_quote":"Mallat's wavelet tour: gives the FDWT/cascade algorithm and orthonormal wavelet basis theory the learnable FDWT layer is built on."},{"cited_title":"Perfect channel splitting by use of interpola- tion/decimation/tree decomposition techniques,","cited_arxiv_id":null,"evidence_quote":"The conjugate quadrature filter concept that yields the g = alternating flip of h constraint and parameter reduction."},{"cited_title":"The wavelet transform, time-frequency localization and signal analysis,","cited_arxiv_id":null,"evidence_quote":"Daubechies wavelets used to initialize the learnable filters and motivate time-frequency localization."},{"cited_title":"A study of stratified random sampling,","cited_arxiv_id":null,"evidence_quote":"Stratified random sampling: the splitting procedure whose speaker-independence is the load-bearing evaluation assumption."}],"review_version":1}