{"id":"c47df0ec-2d72-4dd5-bc72-bea6cb8f05d9","arxiv_id":"2508.19308","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A 1.54M-parameter CNN-RNN with blueprint separable convolutions, time-frequency BiLSTM/LSTM, and spatial/channel attention achieves state-of-the-art infant cry detection on a merged public dataset across SNRs down to -20 dB.","lead":"A lightweight neural network for detecting infant cries in noisy household audio is proposed, combining blueprint separable convolutions with a time-frequency recurrent network. It reports higher accuracy and lower model size than four generic baselines across clean and noisy test conditions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Subject/recording leakage in the 5-fold split may explain the -20 dB gains; no identity-disjoint split is specified.","rationale":"The reader's weakest assumption correctly identifies the most load-bearing threat to the paper's central claim: the evaluation split is not shown to be subject-disjoint, and CryCeleb2023 is an infant-speaker-verification dataset. In good faith, the paper does many things well: it integrates multiple public datasets, applies realistic noise augmentation over a 0 to -20 dB SNR range, reports complexity (1.54M parameters, 0.46 GFLOPs), and provides an ablation. The -20 dB improvements are large and internally coherent. However, the generalization claim for a cry-detection system intended for unseen infants depends on evaluation with unseen speakers/recordings. If the five-fold split is clip-level random, the model can exploit within-infant or within-recording similarity. This is not an accusation of fraud; it is a missing experimental control that the paper does not describe and that the public code can settle. The reader's conditional verdict is appropriate: the concern is significant enough to prevent full acceptance of the SOTA claim, but not severe enough to reject the paper outright, because the issue is testable and fixable. I therefore recommend no change to the reader's CONDITIONAL verdict. The concrete test is straightforward: re-run with a subject/recording-disjoint split and compare the -20 dB numbers.","tokens_in":11095,"tokens_out":3506,"duration_ms":41158,"concrete_test":"Using the released GitHub code, re-run the full Table III protocol with a subject/recording-disjoint 5-fold split: group all 5-second samples by originating infant ID and original recording ID before splitting, so no infant or recording appears in both train and test; use the same augmentation, SNRs, and hyperparameters. Report accuracy/F1 at -20 dB and the per-fold standard deviation. If the proposed method's -20 dB F1 drops below roughly 85% or the gap over ResNet18 narrows to less than 3 pp, the headline robustness claim is not supported; if it stays near 90%, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that at -20 dB the proposed model reaches 91.3% accuracy and 90.8% F1, beating ResNet18 by 6.5 pp accuracy and 9.5 pp F1 (Table III). The validity of this comparison depends entirely on the evaluation protocol. Section IV.A constructs the cry subset from CryCeleb2023, EnesBabyCries1, and iFLYTEK, then splits or concatenates cry segments into 5-second samples. CryCeleb2023 is explicitly a speaker-verification dataset, meaning each infant contributes multiple cry segments; EnesBabyCries1 also contains repeated recordings per infant. Section IV.C states only that 'five-fold cross-validation is performed with the average results reported.' No subject-disjoint or recording-disjoint split is described. If the folds are random at the clip/segment level, the same infant, or even contiguous segments from the same original recording, can appear in both training and test folds. The model can then memorize speaker-specific timbre, pitch, and recording artifacts rather than learn cry-general, noise-robust features. This would inflate clean accuracy (99.0%) and especially the -20 dB numbers, where the model can rely on speaker cues that a deployed system would not have. This concern is load-bearing because it undermines the generalization claim that is the paper's main contribution; it is not merely a missing baseline or a missing variance estimate. The absence of specialized infant-cry baselines is a secondary issue, but even with those baselines, a leaky split would make all comparisons uninterpretable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a lightweight infant cry detection method for noisy environments. The model combines blueprint separable convolutions (BSConv) with multi-scale feature extraction, enhanced spatial attention (ESA), contrast-aware channel attention (CCA), and a time-frequency recurrent module (BiLSTM/LSTM) that the authors call an Adaptive Denoising Module (ADM). Inputs are log Mel-spectrograms extracted from 5-second audio clips. The authors integrate several public datasets (CryCeleb2023, EnesBabyCries1, iFLYTEK, plus non-cry sources) and apply augmentations including speed perturbation, additive noise at SNRs from 0 to -20 dB, reverberation, and time/frequency masking. Experiments compare the method with BiLSTM, ECAPA-TDNN, MobileNetV2, and ResNet18 under clean and noisy conditions, report an ablation study, and give model complexity in parameters and FLOPs. The central empirical claim is that the proposed method achieves 91.3% accuracy and 90.8% F1-score at -20 dB with 1.54M parameters and 0.46G FLOPs, outperforming the compared backbones.","tokens_in":11475,"tokens_out":2966,"duration_ms":32646,"significance":"If the reported results are valid, the paper makes a useful contribution to edge-deployable infant cry detection: the model is small (1.54M parameters), computationally light (0.46G FLOPs), and shows markedly better robustness at -20 dB than the generic backbones. The code is released, the dataset integration across multiple public corpora is a strength, and the T-F recurrent design for noise robustness is a reasonable architectural idea. The ablation study indicates that each component contributes. However, the empirical claims rest entirely on the evaluation protocol, and the protocol has a potentially serious subject-leakage problem that must be resolved before the results can be trusted.","major_comments":[{"comment":"The five-fold cross-validation may suffer from subject/recording leakage. CryCeleb2023 is explicitly a speaker-verification dataset with multiple cry segments per infant, and EnesBabyCries1 also contains repeated recordings per infant. The paper states only that 'five-fold cross-validation is performed with the average results reported' and does not describe a subject-disjoint or recording-disjoint split. If the folds are random at the 5-second clip level, the same infant or contiguous segments from the same original recording can appear in both training and test folds, allowing the model to memorize speaker-specific timbre, pitch, and recording artifacts. This would inflate the reported accuracy, especially at -20 dB (Table III), and undermine the paper's main generalization claim. Please clarify the exact splitting procedure and, ideally, rerun the experiments with a subject-disjoint s","section":"Sec. IV.A and IV.C, Table III"},{"comment":"No variance or confidence intervals are reported. All results are point estimates averaged over five folds, and no standard deviation, per-fold results, or statistical significance tests are given. While the -20 dB margin over ResNet18 (91.3 vs. 84.8 accuracy) is large, some comparisons at 0 dB and in the ablation are close, so without error bars it is difficult to assess whether the claimed improvements are reliable. Please report per-fold results or standard deviations, and where possible a paired significance test.","section":"Sec. IV.C and Tables II-IV"},{"comment":"The comparison set is limited to generic audio/speaker backbones (BiLSTM, ECAPA-TDNN, MobileNetV2, ResNet18). No specialized infant cry detectors from the cited related work (e.g., Refs. [1], [2], [6], [31]) are included, so the statement that the method 'exceeds many state-of-the-art methods' is only supported against four generic models. Adding at least one or two cry-specific baselines, or qualifying the claim, would make the evaluation more convincing.","section":"Sec. IV.C and Table III"}],"minor_comments":[{"comment":"Typo: 'OBATINED' should be 'OBTAINED'.","section":"Table II caption"},{"comment":"The subscript '𝒸' is nonstandard and may render incorrectly; use a plain 'c' consistently. Also, the square-root expression should be checked for clarity: the formula as written uses 'm' both inside and outside the sum, and the reader must infer the intended standard deviation plus mean.","section":"Eq. (4)"},{"comment":"The text says 'convolution groups, implemented by BSConv as well' but the caption of Fig. 1 is not visible in the text; ensure the figure clearly labels BSConv, max-pooling, and bilinear upsampling.","section":"Sec. III.A.2 / Fig. 1"},{"comment":"The phrase 'with 5 seconds in length' is awkward; use 'of 5 seconds' or '5-second recordings'.","section":"Sec. IV.A"},{"comment":"There is a typo in the author list: 'Levr ro' should likely be 'Levrero'.","section":"References [25]"},{"comment":"The speed perturbation description says 'uniformly selects a speed-up factor from 0.8 to 1.2 relative to the original sampling rate' but does not state whether the resampled audio is time-stretched without pitch change or with pitch change. This affects the interpretation of the augmentation.","section":"Sec. IV.B"},{"comment":"The phrase 'exceeds many state-of-the-art methods' is vague; specify the actual comparison set in the abstract or conclusions.","section":"Sec. V / Abstract"},{"comment":"The ADM is called 'Adaptive Denoising Module' but the model is trained only with a classification loss; no denoising reconstruction objective is used. The term 'denoising' is therefore used metaphorically. Consider clarifying that the module learns noise-invariant representations rather than explicitly reconstructing a clean signal.","section":"Sec. III.B"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague:\n\nThe paper is a competent engineering application: it takes BSConv, a time-frequency BiLSTM/LSTM denoising stage, and two attention modules, adds a merged seven-dataset recipe, and reports consistent gains over four generic backbones across SNR levels, with 1.54M parameters and 0.46G FLOPs. The ablation shows the ADM module matters most, and the code is public. That is genuinely useful for people building edge baby monitors.\n\nThe main issue is the evaluation protocol. Five-fold cross-validation is described only as \"average results reported,\" with no mention of a subject-disjoint or recording-disjoint split. CryCeleb2023 is a speaker-verification dataset, so random clip-level folds can place the same infant's cries in both training and test sets. That would let the model memorize voice identity rather than learn cry-general, noise-robust features, and would inflate the clean accuracy (99.0%) and the -20 dB numbers (91.3% accuracy, 90.8% F1). This is not a missing nicety; it is the load-bearing wall for the central claim. I would not call the paper a takedown—the architecture is reasonable and the gains over ResNet18 are plausible even with leakage—but the numbers as reported cannot be taken at face value.\n\nSecondary issues: no variance or error bars anywhere; no comparison with specialized infant-cry detectors, so \"exceeds many SOTA\" is only shown against generic backbones; and hyperparameters (SNR range, mel bands, segment length) are hand-chosen without sensitivity analysis. These are minor relative to the split concern.\n\nWho gets value: practitioners in audio event detection and baby-care products who want a lightweight model and a data-augmentation recipe. The paper is worth a serious referee: it should be reviewed, but the review should require a subject-disjoint split, variance reporting, and at least one specialized baseline. I'd suggest engaging with it as a solid-but-unverified application, not as a SOTA reference.\n\nRecommendation: send it to peer review, but the evaluation protocol needs correction before acceptance.","headline":"A solid engineering combination with plausible gains, but the split protocol must be clarified before trusting the -20 dB claims.","tokens_in":11905,"tokens_out":2289,"would_cite":false,"duration_ms":22773,"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 lightweight cry detector built from blueprint separable convolutions and a time-frequency recurrent module stays accurate down to −20 dB SNR, beating much larger baselines.","keywords":["Infant cry detection","Blueprint separable convolution","Time-frequency recurrent neural network","Noise robustness","Log Mel-spectrogram","Attention mechanism","Low-complexity audio classification","Data augmentation"],"falsifier":"Run a speaker-disjoint evaluation on CryCeleb2023, ensuring no vocalization of any infant in the training set appears in the test set; if the F1-score at −20 dB falls to the level of ResNet18's, the robustness claim fails. A complementary check is deploying the model on continuous, unsegmented home audio and measuring false-alarm rates per hour.","tokens_in":11010,"feed_emoji":"👶","tokens_out":5482,"duration_ms":54439,"temperature":0.7,"pith_summary":"This paper sets out to show that infant cry detection can be made both lightweight and noise-robust enough for home monitoring. It builds a detector from blueprint separable convolutions—which factor each standard filter into a pointwise weight shared across channels and a 2D blueprint—plus a time-frequency recurrent denoising stage and two attention mechanisms. Trained on a multi-source public corpus corrupted with household noise, the model reports 91.3% accuracy and 90.8% F1-score at −20 dB signal-to-noise ratio, above ResNet18, MobileNetV2, ECAPA-TDNN, and BiLSTM, with 1.54 million parameters and 0.46 GFLOPs. The practical payoff, if correct, is a cry detector that fits memory-limited baby-care devices without giving up accuracy in noisy rooms.","feed_headline":"Cry detector holds 91.3% accuracy at −20 dB noise","feed_subtitle":"A 1.54M-parameter audio model outlasts larger baselines in heavy household noise.","key_machinery":"The load-bearing identity is the Blueprint Separable Convolution decomposition, which rewrites the nth filter's c-th channel as ω_c^(n)·B^(n), a per-channel scalar times a shared 2D blueprint. That cuts the parameter count from C·K²·N to (C+K²)·N while keeping most of a regular convolution's expressiveness. The other mechanism is the Adaptive Denoising Module: a BiLSTM processes each STFT frame to capture harmonic structure, an LSTM then tracks temporal dependencies per frequency bin, and residual connections stabilize training; together with contrast-aware channel attention (mean-plus-standard-deviation pooling) and enhanced spatial attention, the model is meant to suppress noise rather tha","core_discovery":"On its own terms, the paper's central claim is that replacing standard convolutions with blueprint separable convolutions, then feeding multi-scale embeddings through a BiLSTM/LSTM adaptive denoising module, gives a cry/non-cry classifier that degrades gracefully as noise rises. When trained on clean audio and tested on 0 dB mixtures, the proposed model loses only 19.0 accuracy points versus 24.3–31.1 for the baselines, keeping F1-score at 74.3% where the next best is 63.9%. When trained on noisy audio covering 0 to −20 dB SNR, the model holds 91.3% accuracy and 90.8% F1-score at the hardest condition, against ResNet18's 84.8% and 81.3%. Its parameter count (1.54M) and FLOPs (0.46G) are smal","pith_inferences":["The five-fold split described in the paper is clip-level, so the same infant's vocalizations can appear in both training and test folds; a speaker-disjoint split would test whether the model learned cry-general features rather than speaker identity.","Because the architecture is generic (BSConv encoder plus a time-frequency recurrent module), the same recipe should transfer to other short acoustic events such as coughs, screams, or pet vocalizations; this extension is not tested in the paper.","The −20 dB condition implies strong denoising; comparing the method against a standard speech-enhancement front end feeding the same classifier would isolate what the ADM contributes.","The paper suggests an energy-threshold sliding-window deployment but does not evaluate it; measuring false-alarm rates on continuous home recordings would confirm the practical claim."],"forward_implications":["A detector with 1.54M parameters and 0.46 GFLOPs fits the memory and power budgets of typical infant-monitor hardware, making real-time on-device cry detection plausible.","Training on noisy, multi-source data is not optional: the paper's clean-trained model loses 25.5 F1 points on 0 dB noise, while the noisy-trained model stays near 98% F1 at 0 dB.","The robustness gain concentrates in the recurrent denoising stage; removing ADM costs about 3.7 F1 points at 0 dB, more than removing ESA, CCA, or multi-scale concatenation.","At −20 dB, the method keeps recall at 91.2%, while every compared baseline falls below 71%; a detector that works at such SNR could catch cries buried under loud household sounds."],"supporting_citations":[{"why":"Introduces Blueprint Separable Convolutions, the parameter-reduction identity the whole encoder rests on.","marker":"[18]"},{"why":"Applies BSConv to low-complexity audio classification, providing prior evidence that the decomposition works for spectrogram-style inputs.","marker":"[19]"},{"why":"Supplies the largest source of infant cry recordings, CryCeleb2023, which anchors the positive-class dataset.","marker":"[24]"},{"why":"ResNet18 is a primary baseline; the paper must beat its accuracy and F1-score while using far fewer parameters.","marker":"[45]"},{"why":"ECAPA-TDNN is a strong audio-classification/speaker-embedding baseline used for comparison in Tables II and III.","marker":"[46]"},{"why":"MobileNetV2 is the lightweight baseline that defines the complexity target the proposed model is compared against.","marker":"[47]"},{"why":"Provides room impulse responses used to add reverberation during training, part of the environmental corruption strategy.","marker":"[44]"},{"why":"Supplies the cyclical learning-rate schedule used to train all models in the comparison.","marker":"[48]"}],"fun_headline_variants":["91.3% infant cry accuracy at −20 dB, 1.54M parameters","Cry detection at −20 dB: 91.3% accuracy from a 1.54M model","Tiny audio net sees through −20 dB noise for cry detection, 91.3%","Cry detector with 1.54M params hits 91.3% at −20 dB noise"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported results rest on five-fold cross-validation over randomly split 5-second clips from public datasets; if clips from the same infant land in both training and test folds, the model could be recalling infant-specific voice traits rather than learning general cry patterns, inflating the numbers relative to unseen-baby deployments.","fun_headline_variants_meta":{"raw":{"variants":["91.3% infant cry accuracy at −20 dB, 1.54M parameters","Cry detection at −20 dB: 91.3% accuracy from a 1.54M model","Tiny audio net sees through −20 dB noise for cry detection, 91.3%","Cry detector with 1.54M params hits 91.3% at −20 dB noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000622,"raw_usage":{"total_tokens":2718,"prompt_tokens":741,"completion_tokens":1977,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1874}},"tokens_in":485,"tokens_out":1977,"duration_ms":18287,"temperature":1.0,"reasoning_tokens":1874,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T16:11:46.770284+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a speaker-disjoint evaluation on CryCeleb2023, ensuring no vocalization of any infant in the training set appears in the test set; if the F1-score at −20 dB falls to the level of ResNet18's, the robustness claim fails. A complementary check is deploying the model on continuous, unsegmented home audio and measuring false-alarm rates per hour.","supporting_citations":[{"cited_title":"Rethinking depthwise separable convo- lutions: How intra-kernel correlations lead to improved mobilenets,","cited_arxiv_id":null,"evidence_quote":"Introduces Blueprint Separable Convolutions, the parameter-reduction identity the whole encoder rests on."},{"cited_title":"Low-complexity acoustic scene classification using blueprint separable convolution and knowledge distillation ,","cited_arxiv_id":null,"evidence_quote":"Applies BSConv to low-complexity audio classification, providing prior evidence that the decomposition works for spectrogram-style inputs."},{"cited_title":"Cryceleb: a speaker verification dataset based on infant cry sounds,","cited_arxiv_id":null,"evidence_quote":"Supplies the largest source of infant cry recordings, CryCeleb2023, which anchors the positive-class dataset."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"ResNet18 is a primary baseline; the paper must beat its accuracy and F1-score while using far fewer parameters."},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks,","cited_arxiv_id":null,"evidence_quote":"MobileNetV2 is the lightweight baseline that defines the complexity target the proposed model is compared against."},{"cited_title":"A study on data augmentation of reverberant speech for robust speech recognition,","cited_arxiv_id":null,"evidence_quote":"Provides room impulse responses used to add reverberation during training, part of the environmental corruption strategy."},{"cited_title":"A cyclical learning rate method in deep learning training,","cited_arxiv_id":null,"evidence_quote":"Supplies the cyclical learning-rate schedule used to train all models in the comparison."}],"review_version":1}