{"id":"89528a89-dcaf-4580-ad31-8e8caa19ebb3","arxiv_id":"2509.02471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ESTM, a dual-branch Mamba with frequency/time patches and a statistical gating module, reports the best average AUC and pAUC on DCASE 2020 Task 2.","lead":"This paper presents ESTM, a neural network for detecting abnormal machine sounds by combining Mamba state-space models with time and frequency separated analysis. It reports top scores on the DCASE 2020 Task 2 benchmark, though the comparisons may not use identical training data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on an uncontrolled, protocol-mismatched comparison with quoted baselines","rationale":"The reader's identified weakest assumption—protocol mismatch due to training on additional data—is plausible and worth testing, but it is not fully certain because DCASE 2020 Task 2 officially provides an additional training dataset and many published baselines also use it. The more robust formulation of the concern is that all baselines are quoted rather than re-run under identical conditions, so the SOTA comparison is uncontrolled. This is a real, load-bearing issue for the central claim, but it does not disprove the result; it only prevents taking the SOTA wording at face value. The internal ablation is consistent with the paper's architectural story, and the method is clearly described. I therefore agree with the original CONDITIONAL verdict but would phrase the condition as 'protocol-controlled re-evaluation of baselines' rather than assuming the additional-data mismatch is necessarily true. No stronger verdict such as REJECT is warranted because the paper's internal evidence and ablations are coherent, and no fatal flaw in the method itself is apparent. A concrete retraining test will settle whether the margin is real.","tokens_in":8102,"tokens_out":5459,"duration_ms":65447,"concrete_test":"Retrain at least two strong baselines—e.g., STgram-MFN [6] and ASD-AFPA [24]—using their released code under the exact ESTM training protocol (dev+additional normal sounds, STFT 1024/hop 512, 128 Mel bins, AdamW lr 1e-4, 200 epochs, same test split) and report mean and std over at least 3 seeds. If ESTM still exceeds both by roughly 1% AUC or more, the protocol-mismatch concern is refuted; if the margin shrinks or reverses, the SOTA claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—ESTM achieves state-of-the-art AUC/pAUC on DCASE 2020 Task 2—depends entirely on Table I, yet the comparison is not protocol-matched. Section III-A states ESTM is trained on normal sounds from both the development and additional datasets, while Section III-C says all baseline entries are 'best reported results from their respective papers.' If those baselines were trained only on the development set, or used different preprocessing/splits, the reported margins (0.84% AUC and 1.79% pAUC; although the numbers in Table I imply 1.85% and 2.54% vs ASD-AFPA) could reflect extra training data rather than the dual-branch Mamba or TSG module. No error bars, significance tests, or repeated-seed results are provided, and no code is released, so the SOTA wording is supported only by this uncontrolled comparison. The internal ablation in Table II does support the contribution of each module within ESTM's own pipeline, but it does not establish superiority over external methods.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ESTM, a dual-branch (time and frequency) Mamba architecture for anomalous sound detection on the DCASE 2020 Task 2 dataset. Inputs are a concatenation of a Log-Mel spectrogram, a TriStat-Gating (TSG) enhanced spectrogram, and the Tgram raw-audio feature. The two Mamba branches scan time and frequency patches separately and are fused before an ArcFace classification head. The authors report the best average AUC (95.76%) and pAUC (90.49%) in Table I and present ablations (Table II) showing that both the TSG module and the dual-path STMamba contribute to performance.","tokens_in":8361,"tokens_out":5136,"duration_ms":55575,"significance":"If the reported results are protocol-matched and reproducible, the paper would present a strong new application of state-space models to ASD, with a plausible architectural justification and consistent internal ablations. The TSG enhancement idea is simple and potentially reusable. Credit is due for the clear ablation design and for evaluating on a public benchmark. However, the central SOTA claim rests on an uncontrolled comparison with published baselines, the results are single-run point estimates, and several numeric statements in the text do not match the tables. These issues materially limit the strength of the contribution as currently written.","major_comments":[{"comment":"The SOTA claim is not protocol-matched. Section III-A states that ESTM trains on normal sounds from the development and additional datasets, while Section III-C says the baselines in Table I are 'best reported results from their respective papers.' It is not reported whether those baselines also used the additional training data. If they did not, the reported margins (e.g., 0.84%/1.79% or the actual Table I differences) could be due to extra training data rather than the architecture. Please retrain baselines under the same protocol, or clearly list which baselines used additional data and explicitly restrict the SOTA wording to the matched setting.","section":"§III-A and §III-C, Table I"},{"comment":"All results are reported as single-run point estimates with no seeds, error bars, or significance tests. Since several improvements are small (e.g., the average AUC gain of the TSG module in Table II is about 0.9 percentage points), the reader cannot assess whether these are stable effects or training noise. Please provide multiple runs (at least 3–5 seeds), report mean±std, and ideally a paired test across machine types.","section":"Section III-B, Table I, Table II"},{"comment":"There are multiple numerical inconsistencies. The abstract and §III-C claim an improvement of 0.84% AUC and 1.79% pAUC over ASD-AFPA, but Table I implies 95.76−93.91=1.85 and 90.49−87.95=2.54. In §III-D, the text says STMamba with ESTgram improves average AUC by 0.91% over STgram, but Table II gives 95.76−94.90=0.86; the stated 2.07%/1.94% gains over SMamba/TMamba also do not match the table means (differences are 1.94/1.82). The ToyConveyor gains in §III-C (7.99%/4.50%) do not correspond to the values in Table I either. These numbers need to be recalculated and corrected, and the tables/text must be consistent.","section":"Abstract, §III-C, §III-D, Table I, Table II"},{"comment":"The evaluation split is described ambiguously. Section III-A says 'the test data ... from the development dataset' was used, but Table I's caption reads 'on the test data of the development and additional datasets.' This matters for baseline comparability and for reproducing the results. Please specify exactly which test folds were used (development, additional, or both) and confirm that the quoted baselines used the same folds.","section":"§III-A, Table I caption"}],"minor_comments":[{"comment":"The TSG module is described as 'parameter-free,' but Eq. (1) includes a scaling factor α that is manually set to 2 in §III-B. This should be rephrased, e.g., 'with a single scalar hyperparameter,' to avoid contradiction.","section":"Eq. (1) and §I contributions"},{"comment":"In Algorithm 1, a Conv2D operation is applied to the token sequence. Since the input is a 1D token sequence, the distinction between Conv2D and a 1D/2D convolution should be clarified for reproducibility.","section":"Algorithm 1"},{"comment":"The notation W is used for the projection matrix and also for the frequency window size in the preceding paragraph; please disambiguate.","section":"Section II-B, Eq. (4)–(5)"},{"comment":"The phrase 'state-of-the-art ASD-AFPA method' is odd because ASD-AFPA is a baseline in Table I that is outperformed by TASTgram(NAMix) in average AUC. The text should state the best baseline actually used for comparison.","section":"Section III-C"},{"comment":"No code is released. Providing code or detailed hyperparameter/config files would substantially aid reproducibility, especially given the single-run results.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The core architecture and ablation are reasonable, and the paper is within scope for a signal-processing/journal venue. The main risk is the uncontrolled SOTA comparison: the authors use the additional training set without matching baselines. If the authors can provide a matched comparison or reframe claims as 'improvement over our own baselines,' the paper would be acceptable. Also, the numeric inconsistencies are pervasive enough that I recommend a careful audit before the next submission. I have no concerns about novelty disclosure or citation practices beyond the protocol issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: ESTM is a cleanly-designed dual-branch Mamba for industrial ASD, and the internal ablation is the most convincing part of the paper. The headline SOTA claim, however, is built on an uncontrolled comparison and a small internal inconsistency, so I wouldn't take the margin at face value yet.\n\nWhat's new: the combination of time/frequency patchified Mamba branches with the TriStat-Gating module (median/RMS/variance on Mel frames) is a legitimate new architecture. The paper is clearly written, the math is straightforward, and there is no circularity. Table II is the strongest evidence: swapping STgram for ESTgram improves every machine type, and the dual-path STMamba beats either branch alone by about 2% AUC under the same input. That's a coherent, if modest, story.\n\nSoft spots: (1) The central Table I comparison is not apples-to-apples. Section III-A says training uses normal sounds from both development and additional datasets, while the baselines are quoted as best-reported results from their respective papers, which typically train only on development. The 0.84% average AUC edge (or 1.85%, depending on which column you read—the text and Table I disagree) could partly or fully reflect extra training data. No retrained baselines, no error bars, no repeated seeds, and no code are provided. (2) The stated improvement of 0.84% AUC and 1.79% pAUC does not match the numbers in Table I; the actual margins versus ASD-AFPA are 1.85% and 2.54%. That's a small but telling sloppiness. (3) The 'first decoupled dual-path Mamba for ASD' claim is a bit overstated given existing dual-path work (CLP-SCF) and audio Mamba applications, but that's minor.\n\nBottom line: I'd send this to peer review. The architecture is plausible and the ablation is internally consistent; a referee should ask for protocol-matched baselines, variance estimates, and code before the SOTA claim is accepted. It's also a useful reading-group example of how training-set mismatches can inflate benchmark claims.","headline":"ESTM is a well-built dual-branch Mamba with a clean internal ablation, but its SOTA margin rests on a protocol-mismatched comparison and needs verification.","tokens_in":8819,"tokens_out":2596,"would_cite":true,"duration_ms":28633,"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":"Dual-branch Mamba tops machine-anomaly sound detection at 95.76 AUC","keywords":["anomalous sound detection","Mamba","selective state-space model","spectral-temporal fusion","self-supervised classification","Mel spectrogram enhancement","DCASE 2020 Task 2","time-frequency decoupling"],"falsifier":"Take the six methods in Table I and retrain each on the same normal-sound training set used for ESTM (development plus additional), then evaluate on the same development test split. If ESTM no longer holds at least a 0.84-point average-AUC lead, the headline claim is refuted. A second check: train ESTM on the development set only; if its average AUC drops near or below the published baselines, most of the advantage comes from extra training data rather than the architecture.","tokens_in":8023,"feed_emoji":"🔊","tokens_out":8936,"duration_ms":93699,"temperature":0.7,"pith_summary":"This paper sets out to show that anomalous machine sounds can be detected more reliably when the detector models time and frequency separately with selective state-space (Mamba) layers instead of relying on local receptive fields. Its ESTM system feeds a fused spectrogram—raw Log-Mel, a statistically gated Mel version, and the raw-waveform Tgram—into two patch-based Mamba branches, one scanning along frequency and one along time, then trains the whole network as an ID classifier with ArcFace loss. On the DCASE 2020 Task 2 benchmark, ESTM is reported to reach an average AUC of 95.76% and pAUC of 90.49%, exceeding the best prior published average by 0.84 and 1.79 points. The significance for a general reader: this is a concrete step toward recognizing never-before-heard machine failures from normal-only recordings, a setting where long-term trends and cross-band energy shifts matter.","feed_headline":"Dual-branch Mamba tops machine-anomaly sound detection at 95.76 AUC","feed_subtitle":"A time-frequency decoupled Mamba with statistically gated spectrograms beats prior ASD methods by 0.84% AUC.","key_machinery":"The STMamba backbone is the central mechanism: two selective-state-space (Mamba) branches, SMamba and TMamba, scan patched spectrograms along the frequency axis and the time axis respectively, each using learnable class tokens and positional embeddings, then align and add their outputs. The supporting mechanism is the TriStat-Gating (TSG) module, which multiplies the Log-Mel spectrogram frame-wise by a sigmoid-gated combination of median, RMS, and variance statistics to produce the enhanced ESgram. Together they are meant to capture long-range temporal dependencies and cross-band energy coupling that local-receptive-field models miss.","core_discovery":"The paper argues that anomalous sound detection improves when a spectrogram is processed by two separate selective state-space branches, one scanning frequency patches and the other scanning time patches, and their outputs are fused. Before the scan, the Log-Mel spectrogram is enhanced by a parameter-free TriStat-Gating module that multiplies each time frame by a sigmoid-gated combination of median, RMS, and variance; this gated spectrogram is concatenated with the original Log-Mel and the raw-signal Tgram. The network is trained as a self-supervised machine-ID classifier using ArcFace loss, and the negative log probability serves as the anomaly score. On the DCASE 2020 Task 2 development te","pith_inferences":["Inference: The headline margin depends on training ESTM on normal sounds from both the development and additional sets, while the quoted baselines are 'best reported results from their respective papers'; a matched-data retraining is the test that would confirm or erase the 0.84/1.79-point lead.","Inference: The TSG gating is parameter-free and computed frame-wise from three statistics, so it could be lifted out of ESTM and applied to other spectrogram-based ASD models as a preprocessing step; the paper only tests it inside its own backbone.","Inference: The ablation section's explanation for ToyConveyor—that temporal fusion paradoxically dilutes frequency-domain features—is contradicted by Table II, where the fused STMamba outperforms both single branches on that machine type; a reader should not carry that explanation forward.","Inference: On the newer DCASE ASD tasks that add domain shift, ESTM's long-range temporal branch may show different gains; the paper evaluates only the 2020 Task 2 split, so transfer is untested."],"forward_implications":["If ESTM's numbers hold, the best published anomalous-sound-detection scores on the DCASE 2020 Task 2 development test set move from transformer/CNN hybrids to a dual-branch selective state-space architecture.","The TSG enhancement is parameter-free and consistently improves AUC across all machine types; the full model gains 0.91% average AUC over the same backbone with the un-enhanced input.","The time-frequency decoupled model outperforms either single branch by roughly 2 AUC points, supporting the paper's claim that temporal and spectral branches are complementary.","The ToyConveyor case shows the largest per-machine improvement, indicating that the method is especially strong on spectrally dominant anomaly patterns."],"supporting_citations":[{"why":"Supplies the self-supervised ID-classification training scheme and the Tgram raw-audio feature that ESTM extends.","marker":"[6]"},{"why":"The ASD-AFPA system whose reported average AUC/pAUC is the baseline ESTM claims to beat by 0.84 and 1.79 points.","marker":"[24]"},{"why":"Provides the selective state-space model that both branches of STMamba use for long-range sequence modeling.","marker":"[14]"},{"why":"Contributes the patch-embedding plus learnable-class-token formulation ESTM adapts for time and frequency tokens.","marker":"[25]"},{"why":"ArcFace loss that tightens intra-class distributions and increases inter-class separation in the ID-classification head.","marker":"[26]"},{"why":"MIMII dataset supplying real machine operating sounds for the DCASE 2020 Task 2 benchmark.","marker":"[27]"},{"why":"ToyADMOS dataset supplying the synthetic miniature-machine sounds used alongside MIMII in the benchmark.","marker":"[28]"}],"fun_headline_variants":["Mamba dual-branch beats baselines in sound anomaly detection","Time-frequency Mamba improves machine anomaly detection AUC","Dual-path Mamba with gated spectrogram tops sound fault detection","State-space dual branch sharpens anomaly detection in machines","Mamba's time-frequency split nudges ASD accuracy up 0.84%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The headline claim is load-bearing on the assumption that the quoted baseline numbers are directly comparable to ESTM, even though ESTM was trained on normal sounds from both the development and additional datasets while the baselines were not.","fun_headline_variants_meta":{"raw":{"variants":["Mamba dual-branch beats baselines in sound anomaly detection","Time-frequency Mamba improves machine anomaly detection AUC","Dual-path Mamba with gated spectrogram tops sound fault detection","State-space dual branch sharpens anomaly detection in machines","Mamba's time-frequency split nudges ASD accuracy up 0.84%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001055,"raw_usage":{"total_tokens":4242,"prompt_tokens":698,"completion_tokens":3544,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":3471}},"tokens_in":442,"tokens_out":3544,"duration_ms":26290,"temperature":1.0,"reasoning_tokens":3471,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:34:34.419558+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the six methods in Table I and retrain each on the same normal-sound training set used for ESTM (development plus additional), then evaluate on the same development test split. If ESTM no longer holds at least a 0.84-point average-AUC lead, the headline claim is refuted. A second check: train ESTM on the development set only; if its average AUC drops near or below the published baselines, most of the advantage comes from extra training data rather than the architecture.","supporting_citations":[{"cited_title":"Anomalous sound detection using spectral-temporal information fusion,","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised ID-classification training scheme and the Tgram raw-audio feature that ESTM extends."},{"cited_title":"Anomalous sound detection using self-attention-based frequency pattern analysis of machine sounds,","cited_arxiv_id":null,"evidence_quote":"The ASD-AFPA system whose reported average AUC/pAUC is the baseline ESTM claims to beat by 0.84 and 1.79 points."},{"cited_title":"Vision mamba: Efficient visual representation learning with bidirectional state space model,","cited_arxiv_id":null,"evidence_quote":"Contributes the patch-embedding plus learnable-class-token formulation ESTM adapts for time and frequency tokens."},{"cited_title":"Arcface: Additive angular margin loss for deep face recognition,","cited_arxiv_id":null,"evidence_quote":"ArcFace loss that tightens intra-class distributions and increases inter-class separation in the ID-classification head."},{"cited_title":"MIMII Dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,","cited_arxiv_id":null,"evidence_quote":"MIMII dataset supplying real machine operating sounds for the DCASE 2020 Task 2 benchmark."},{"cited_title":"ToyADMOS: A dataset of miniature-machine operating sounds for anomalous sound detection,","cited_arxiv_id":null,"evidence_quote":"ToyADMOS dataset supplying the synthetic miniature-machine sounds used alongside MIMII in the benchmark."}],"review_version":1}