Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Mind the Gap: Detecting Cluster Exits for Robust Local Density-Based Score Normalization in Anomalous Sound Detection

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Local density-based score normalization fails when a fixed neighborhood crosses a cluster boundary; a per-sample jump detector that truncates at the first cluster exit restores stability across neighborhood sizes.

desk verdict The adaptive-neighborhood idea is promising and the evaluation is solid, but the algorithm's '+1' means it includes the first neighbor beyond the detected gap, so the central locality-preservation claim is not supported as written. read the letter →

arxiv 2602.18777 v2 pith:QD3ZUFY6 submitted 2026-02-21 eess.AS cs.SD

Kevin Wilkinghoff , Gordon Wichern , Jonathan Le Roux , Zheng-Hua Tan This is my paper
classification eess.AScs.SD
keywords anomaloussounddetectionscorenormalizationlocaldensityestimationneighborhoodsizeclusterexitdomainshiftembeddingspacetraining-free
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the performance collapse of local density-based score normalization (LDN) as the neighborhood size grows is structural, not statistical: the neighborhood crosses cluster boundaries in the embedding space, violating the locality assumption behind density estimation. To counter this, the authors propose cluster exit detection (CED), a training-free, per-sample rule that looks for sharp jumps in sorted neighbor distances and truncates the neighborhood at the first jump. Across five embedding models and five public machine-sound benchmarks, CED keeps performance nearly flat over a wide range of neighborhood sizes and yields small but consistent gains over fixed small-neighborhood baselines, mainly on domain-shifted data. The absolute gains are modest, but the message is that the degradation is diagnosable and avoidable without adding a tunable hyperparameter.

What carries the argument

The machinery is the distance-ratio sequence: for each reference sample, r_k = d_k / d_{k+1} for consecutive sorted neighbor distances, values near 0 indicate a sharp jump. Adjacent ratios are averaged to form a smoothed sequence, and the smallest smoothed ratio identifies the most likely cluster exit. A conservative threshold (the 4th percentile of that sequence) decides where expansion becomes unreliable, and a fallback triggers two-neighbor estimation in sparse regions. This per-sample adaptive neighborhood size is then plugged into LDN's local density estimate, making the whole score normalization training-free and parameter-light.

What would settle it

Take an embedding whose reference set has no cluster structure, e.g., uniform samples in a hypercube. With no cluster exits, the smoothed ratio sequence should contain no values below the 4th-percentile threshold, so CED should fall back to the full neighborhood and match the fixed-large-K baseline. If CED still truncates or changes scores, the cluster-exit explanation fails. On data with known cluster memberships, a direct check is whether the per-sample K equals the distance to the true cluster boundary plus one.

Watch

Extended reading notes

Core claim

The central discovery is that the locality assumption of LDN fails at a characteristic scale. For each reference sample, sorted distances to its neighbors grow smoothly inside a cluster and then jump; the jump, called a cluster exit, marks the point where adding further neighbors corrupts the local density estimate. CED detects this exit per sample by locating the smallest smoothed ratio of consecutive distances, truncates the neighborhood at the earliest candidate below the 4th percentile of the ratio sequence, and falls back to two neighbors in sparse regions. Replacing the fixed neighborhood size K with this per-sample estimate changes the failure profile: performance no longer decays as

Load-bearing premise

The load-bearing premise is that the fixed thresholds in the cluster-exit rule transfer across all datasets and embedding models without per-dataset tuning; if these values were chosen with knowledge of the evaluation data, the training-free claim weakens and the reported gains could shrink on unseen benchmarks.

Editorial extensions

If this is right

  • Neighborhood-size sensitivity in LDN is not intrinsic; a per-sample adaptive rule can replace manual tuning of K.
  • Adaptive truncation helps most under domain shift, where cluster exits occur earlier in the target domain.
  • Large fixed neighborhoods do not deliver the gains; the benefit comes from truncation, not from using more neighbors.
  • The rule can be dropped into existing LDN backends with no training and no added inference-time cost, because the normalization constants depend only on the reference set.
  • The principle generalizes: other cluster-exit detectors could replace the ratio heuristic within the same adaptive-normalization scheme.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the mechanism is label-free and training-free, an extension the authors leave implicit is that it should transfer to other distance-based density estimators and outlier detectors, not just LDN, whenever embeddings contain detectable gaps.
  • The fixed thresholds (4th percentile, 0.85, 1.02) are the least principled component; a data-driven jump test or an automatic percentile selection could extend the method to embeddings with very different density profiles.
  • A sharper evaluation than average AUC would measure whether the per-sample neighborhood size coincides with the true cluster boundary on data with known attributes; that would directly confirm the claimed mechanism.
  • On homogeneous single-domain data the method shows no gain, suggesting a gating rule could switch CED off when no distance jumps are present, saving computation and avoiding accidental truncation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes cluster exit detection (CED), a training-free method to adapt the neighborhood size K in local density-based anomaly score normalization (LDN). CED uses distance ratios r_k = d_k/d_{k+1} to detect gaps in the distance profile, selects a per-reference-sample K, and applies LDN/VarMin with this adaptive neighborhood. Experiments on five DCASE benchmark datasets with five embedding models and two LDN variants report reduced sensitivity to K (Fig. 2) and small but consistent average gains (Tables 1-2). The authors interpret these results as evidence that LDN degradation at large K is caused by crossing cluster boundaries, and that CED preserves locality.

Significance. If the mechanism is correct, CED would be an attractive drop-in for LDN: it is simple, has no learned parameters, and removes a critical hyperparameter. The paper's empirical scope is substantial—five datasets, five embeddings, bootstrap CIs, and a systematic K sweep. The main weaknesses are that (i) the adaptive rule appears to include the first out-of-cluster neighbor rather than truncating before it, which breaks the causal link to locality preservation, and (ii) the fixed thresholds are not subjected to sensitivity analysis, leaving a potential overfitting-to-benchmarks concern. The approach merits publication once these issues are resolved.

major comments (3)
  1. [Section 4.2, Eq. (10); Section 4.3, Eq. (11)] The off-by-one in the truncation rule is load-bearing. Since r_k = d_k/d_{k+1}, a small r_k means the gap is between the k-th and (k+1)-th neighbors; a locality-preserving neighborhood includes neighbors 1..k, not 1..k+1. Eq. (10) sets K_hat(y) = min(k_ext, k_min) + 1, which deliberately includes the first neighbor beyond the detected gap. The sentence 'Since ratios require neighbor pairs, the index is effectively shifted' is not a valid justification: d_{k+1} is needed only for detection, and it is not part of the local neighborhood if the gap is at k. The same issue appears in Eq. (11): when r_1 is the small ratio, the fallback uses K=2, although the exit occurs between d_1 and d_2. Thus the algorithm as written is not 'truncating at the first exit'; it truncates one neighbor later. Please revise the algorithm (e.g., K_hat = min(k_ext,k_min), with the corresponding fallback K=1) and re
  2. [Sections 4.2-4.3, Eqs. (8) and (11)] The fixed thresholds (q=0.04, r1<0.85, r1/rmin>1.02) are stated as 'conservative' and 'data-adaptive' only in the sense of using percentiles, but the actual constants are hard-coded without any sensitivity analysis. The central claim that CED requires 'no dataset-specific tuning' and that 'effectiveness stems from detecting distance jumps rather than precise threshold values' is an empirical assertion that needs support. Please provide a sensitivity study (e.g., vary q over {0.01,0.02,0.05,0.1} and the sparse-region thresholds over a grid) across the five datasets and at least a subset of embeddings. If the reported gains are confined to narrow threshold ranges or to values chosen on the test sets, the 'training-free' claim must be qualified.
  3. [Table 1 / Fig. 2] In Table 1, LDN and LDN+VarMin are evaluated at K=2, while LDN+CED and LDN+VarMin+CED are evaluated at K=64. This is not a matched comparison: K is a hyperparameter of the proposed method, and K=64 is chosen because CED is stable there. Since the abstract and Section 3 claim 'consistent performance gains over fixed small-neighborhood baselines,' the comparison should at least include matched-K results (e.g., both at K=2, K=16, K=64) or a clear justification for choosing the best operating point for each method. Fig. 2 partially addresses this, but Table 1, as the primary quantitative summary, is presented in a way that is favorable to CED.
minor comments (4)
  1. [Fig. 2] The x-axis label 'K' is missing in the first three subplots but appears in later ones. Please add the label consistently to all panels.
  2. [Table 1] There are formatting issues with duplicated percentages (e.g., '69.25%69.25%69.25%'). Please clean up the table formatting.
  3. [Reproducibility] The paper does not mention code availability or random seeds for the Direct-ACT ten trials. Providing these would improve reproducibility.
  4. [Section 3, Fig. 1] The causal interpretation that detected distance jumps correspond to cluster boundaries is supported only by a qualitative figure. A quantitative validation (e.g., using machine-type labels or a cluster-quality measure) would strengthen the claim, though it is not essential for the empirical contribution.

Circularity Check

0 steps flagged · score 2.0 of 10

No definitional circularity; minor self-citations and two non-circular robustness/correctness concerns.

full rationale

The paper does not derive its performance claims from a fitted constant or from an equation that is equivalent to its input. CED is a new, training-free rule evaluated on five external DCASE benchmarks, so the AUC comparisons are not a fitted-input-called-prediction reduction. The LDN and VarMin baselines are prior work, including same-author references [18,19,22], but they are used as comparators rather than as forced premises; no uniqueness theorem or ansatz is imported through self-citation. The only potentially load-bearing self-citation is [17] (same authors, 'submitted to TASLP') for the generalized-mean pooling hyperparameters in the experimental setup; this is a preprocessing detail and does not by construction produce the reported gains. I flag two non-circular concerns. First, the fixed thresholds in Eqs. (8) and (11) (Q0.04, 0.85, 1.02) are asserted with no sensitivity analysis, weakening the 'no dataset-specific tuning' claim; this is a reproducibility/data-leakage risk, not a circularity. Second, the skeptic's Eq. (10) off-by-one is real: since r_k=d_k/d_{k+1}, a small r_k is a gap between neighbors k and k+1, so stopping at the exit would imply K_hat=kmin or kext, not kmin+1; the sentence 'Since ratios require neighbor pairs, the index is effectively shifted' does not repair this. That undermines the causal narrative that CED truncates before the cluster exit, but it does not make the empirical comparison circular. Overall, the central derivation is self-contained enough that no circular step can be exhibited; the score reflects minor self-citation and unresolved robustness concerns rather than actual circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

No new physical entities are postulated. The uncharged contributions are the heuristic thresholds and the causal interpretation of distance jumps; the method itself is a deterministic function of distances and fixed constants.

free parameters (5)
  • epsilon (distance-ratio smoothing) = 1e-12
    Added to the denominator in Eq. (5) to avoid division by zero; not materially fitted, but a free constant chosen by the authors.
  • sparse-region fallback threshold r1 < 0.85 = 0.85
    Eq. (11): if the first distance ratio is below 0.85, the reference point is treated as sparse and the neighborhood falls back to two. Hand-set without reported sensitivity analysis.
  • sparse-region ratio threshold r1/rmin > 1.02 = 1.02
    Eq. (11): a very strong ratio drop triggers fallback to two neighbors. Hand-set without reported sensitivity analysis.
  • cluster-exit percentile q = 0.04 = 0.04
    Eq. (8): ratios below the 4th percentile of the per-sample ratio sequence are considered cluster exits. Hand-set without reported sensitivity analysis.
  • maximum candidate neighborhood K for main results = 64
    Table 1 reports CED with K=64 while baselines use K=2; the choice of 64 as the candidate pool affects the comparison and is not justified beyond Fig. 2's stability plot.
assumptions (4)
  • domain assumption Cluster exits can be reliably identified by sharp drops in the smoothed distance-ratio sequence.
    Section 4.2: the entire CED mechanism rests on the belief that a low ~r_k corresponds to crossing a cluster boundary rather than noise or duplicate embeddings.
  • domain assumption Truncating the neighborhood at the first detected cluster exit improves LDN score normalization.
    Section 4.4: the adaptive neighborhood ^K(y) is fed into LDN; this presumes local density estimates are more accurate when truncated at the first distance jump.
  • ad hoc to paper The fixed thresholds (0.04 percentile, 0.85, 1.02) generalize across datasets and embeddings without per-dataset tuning.
    Section 4.2-4.3 and Section 4 text claim 'fixed thresholds are used' and 'stable behavior'; no sensitivity analysis or selection method is provided.
  • domain assumption The performance degradation of LDN with increasing K is caused by crossing cluster boundaries, not by increased estimator variance or other effects.
    Section 3 and Fig. 1: the causal story is supported only by qualitative inspection of average distance profiles, not by a controlled statistical test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Mind the Gap: Detecting Cluster Exits for Robust Local Density-Based Score Normalization in Anomalous Sound Detection." pith.science (2026). https://pith.science/paper/QD3ZUFY6

@misc{pith2026260218777,
  author       = {Pith},
  title        = {Pith review of: Mind the Gap: Detecting Cluster Exits for Robust Local Density-Based Score Normalization in Anomalous Sound Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QD3ZUFY6}},
  note         = {Machine review of arXiv:2602.18777}
}
read the original abstract

Local density-based score normalization is an effective component of distance-based embedding methods for anomalous sound detection, particularly when data densities vary across conditions or domains. In practice, however, performance depends strongly on neighborhood size. Increasing it can degrade detection accuracy when neighborhood expansion crosses cluster boundaries, violating the locality assumption of local density estimation. This observation motivates adapting the neighborhood size based on locality preservation rather than fixing it in advance. We realize this by proposing cluster exit detection, a lightweight mechanism that identifies distance discontinuities and selects neighborhood sizes accordingly. Experiments across multiple embedding models and datasets show improved robustness to neighborhood-size selection and consistent performance gains.

Figures

Figures reproduced from arXiv: 2602.18777 by the authors.

Figure 1
Figure 1. Average sorted distances (left) and distance ratios (right) for BEATs embeddings of the “ToyCar” machine on the DCASE2025 dataset in the source and target domains. Pronounced distance jumps and low ratios mark cluster exits, which occur earlier in the target domain and reveal violations of locality under fixed neighborhood sizes. 2.1. Base anomaly scoring Let Xtest ⊂ R D denote the set of test samples and Xref ⊂ R D… view at source ↗
Figure 2
Figure 2. Performance ratios relative to not applying LDN for ratio-based LDN variants as a function of the neighborhood size K. Values are geometric means across all five evaluated datasets. For Direct-ACT, results are averaged over ten independent trials. to reduce sensitivity to isolated fluctuations. For K = 2, no smoothing is applied. 4.2. Detecting cluster exits As the neighborhood size k increases, a cluster exit is re… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection

    cs.SD 2026-07 conditional novelty 6.0 of 10

    A label-free domain-balance criterion with per-domain quantile calibration rank-predicts DCASE 2025 evaluation scores (ρ=+0.91) where development scores fail, but the transfer is robust in only one of three challenge years.

Reference graph

Works this paper leans on

47 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Introduction Semi-supervised anomalous sound detection (ASD) for ma- chine condition monitoring aims to detect anomalous machine sounds using training data that contain only recordings of nor- mal operation. In practice, systems are often deployed under domain shifts, such as changes in operating conditions, back- ground noise, recording devices, or envir...

  2. [2]

    The key idea is to normalize anomaly scores using statistics derived from a lo- cal neighborhood in the embedding space, thereby accounting for variations in local data density

    Local density-based anomaly score normalization LDN has been shown to improve robustness to domain shifts in distance-based anomaly detection [18, 19]. The key idea is to normalize anomaly scores using statistics derived from a lo- cal neighborhood in the embedding space, thereby accounting for variations in local data density. More recently, VarMin was p...

    arXiv 2026

  3. [3]

    Increasing the neighbor- hood sizeKtherefore assumes that this locality remains valid as the neighborhood expands

    Locality violations and cluster exits LDN assumes that nearby reference samples belong to the same local region in the embedding space. Increasing the neighbor- hood sizeKtherefore assumes that this locality remains valid as the neighborhood expands. In practice, LDN is applied with very small neighborhoods, since increasingKbeyond one or two neighbors of...

  4. [4]

    Cluster exit detection We propose a training-free CED mechanism for adaptive neigh- borhood selection in local density-based normalization. Be- cause local density estimates in LDN are defined with respect to reference neighborhoods, all computations are carried out independently for each reference sampley∈ X ref using only distances to other reference sa...

  5. [5]

    Datasets We evaluate performance on five publicly available datasets for semi-supervised acoustic anomaly detection

    Experimental setup 5.1. Datasets We evaluate performance on five publicly available datasets for semi-supervised acoustic anomaly detection. We consider the DCASE2020 dataset [23], constructed from MIMII [24] and ToyADMOS [25]; DCASE2022 [1], based on MIMII-DG [26] and ToyADMOS2 [27]; DCASE2023 [2], extending MIMII- DG and ToyADMOS2+ [28]; DCASE2024 [3], ...

  6. [6]

    Sensitivity to neighborhood size We first analyze the sensitivity of existing approaches, namely LDN and LDN+VarMin, to the choice of the neighborhood size

    Results and discussion 6.1. Sensitivity to neighborhood size We first analyze the sensitivity of existing approaches, namely LDN and LDN+VarMin, to the choice of the neighborhood size. As shown in Fig. 2, average performance improves only when increasing the neighborhood sizeKfrom1to2. This observa- tion is consistent with the findings in [19], which reco...

    2025

  7. [7]

    Based on this insight, we proposed CED, a lightweight mechanism that detects distance discontinuities to identify neighborhood exits and adapt score normalization ac- cordingly

    Conclusion In this work, we analyzed the sensitivity of LDN to neigh- borhood size and identified a structural failure mode that oc- curs when neighborhood expansion crosses cluster boundaries, thereby violating the locality assumption underlying density estimation. Based on this insight, we proposed CED, a lightweight mechanism that detects distance disc...

  8. [8]

    All scientific content, interpretations, and conclusions are the responsibility of the authors

    Generative AI disclosure Generative AI tools were used for language editing and polish- ing of the manuscript. All scientific content, interpretations, and conclusions are the responsibility of the authors

Show all 47 references
  1. [9]

    Description and discussion on DCASE 2022 Chal- lenge Task 2: Unsupervised anomalous sound detection for ma- chine condition monitoring applying domain generalization tech- niques,

    K. Dohiet al., “Description and discussion on DCASE 2022 Chal- lenge Task 2: Unsupervised anomalous sound detection for ma- chine condition monitoring applying domain generalization tech- niques,” inProc. DCASE, 2022

    2022

  2. [10]

    Description and discussion on DCASE 2023 Challenge Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    ——, “Description and discussion on DCASE 2023 Challenge Task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2023

    2023

  3. [11]

    Description and discussion on DCASE 2024 Challenge Task 2: First-shot unsupervised anomalous sound de- tection for machine condition monitoring,

    T. Nishidaet al., “Description and discussion on DCASE 2024 Challenge Task 2: First-shot unsupervised anomalous sound de- tection for machine condition monitoring,” inProc. DCASE, 2024

    2024

  4. [12]

    Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,

    ——, “Description and discussion on DCASE 2025 challenge task 2: First-shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2025

    2025

  5. [13]

    Self-supervised classification for detecting anomalous sounds,

    R. Giri, S. V . Tenneti, F. Cheng, K. Helwani, U. Isik, and A. Krish- naswamy, “Self-supervised classification for detecting anomalous sounds,” inProc. DCASE, 2020

    2020

  6. [14]

    Anomalous sound detection as a simple binary classification problem with careful selection of proxy outlier examples,

    P. Primus, V . Haunschmid, P. Praher, and G. Widmer, “Anomalous sound detection as a simple binary classification problem with careful selection of proxy outlier examples,” inProc. DCASE, 2020

    2020

  7. [15]

    Sub-cluster AdaCos: Learning representations for anomalous sound detection,

    K. Wilkinghoff, “Sub-cluster AdaCos: Learning representations for anomalous sound detection,” inProc. IJCNN, 2021

    2021

  8. [16]

    Improved domain generalization via disentangled multi-task learning in unsupervised anomalous sound detection,

    S. Venkatesh, G. Wichern, A. S. Subramanian, and J. Le Roux, “Improved domain generalization via disentangled multi-task learning in unsupervised anomalous sound detection,” inProc. DCASE, 2022

    2022

  9. [17]

    AnoPatch: Towards better consistency in machine anomalous sound detection,

    A. Jianget al., “AnoPatch: Towards better consistency in machine anomalous sound detection,” inProc. Interspeech, 2024

    2024

  10. [18]

    Adaptive prototype learning for anomalous sound detec- tion with partially known attributes,

    ——, “Adaptive prototype learning for anomalous sound detec- tion with partially known attributes,” inProc. ICASSP, 2025

    2025

  11. [19]

    Zero- shot anomalous sound detection in domestic environments using large-scale pretrained audio pattern recognition models,

    A. I. Mezza, G. Zanetti, M. Cobos, and F. Antonacci, “Zero- shot anomalous sound detection in domestic environments using large-scale pretrained audio pattern recognition models,” inProc. ICASSP, 2023

    2023

  12. [20]

    Deep generic representations for domain-generalized anomalous sound detection,

    P. Saengthong and T. Shinozaki, “Deep generic representations for domain-generalized anomalous sound detection,” inProc. ICASSP, 2025

    2025

  13. [21]

    Towards few-shot training-free anomaly sound detection,

    H.-H. Wu, W.-C. Lin, A. Kumar, L. Bondi, S. Ghaffarzadegan, and J. P. Bello, “Towards few-shot training-free anomaly sound detection,” inProc. Interspeech, 2025

    2025

  14. [22]

    Exploring self-supervised audio models for generalized anomalous sound detection,

    B. Han, A. Jiang, X. Zheng, W.-Q. Zhang, J. Liu, P. Fan, and Y . Qian, “Exploring self-supervised audio models for generalized anomalous sound detection,”IEEE Trans. Audio, Speech, Lang. Process., vol. 33, 2025

    2025

  15. [23]

    ECHO: Frequency-aware hi- erarchical encoding for variable-length signal,

    Y . Zhang, J. Liu, and M. Li, “ECHO: Frequency-aware hi- erarchical encoding for variable-length signal,”arXiv preprint arXiv:2508.14689, 2025

    2025

  16. [24]

    FISHER: A foundation model for multi-modal industrial signal comprehensive representation,

    P. Fanet al., “FISHER: A foundation model for multi-modal industrial signal comprehensive representation,”arXiv preprint arXiv:2507.16696, 2025

    2025 arXiv

  17. [25]

    Temporal pooling strategies for training-free anomalous sound detection with self- supervised audio embeddings,

    K. Wilkinghoff, S. Yadav, and Z.-H. Tan, “Temporal pooling strategies for training-free anomalous sound detection with self- supervised audio embeddings,” 2026, submitted to TASLP

    2026

  18. [26]

    Keeping the balance: Anomaly score calculation for domain generalization,

    K. Wilkinghoff, H. Yang, J. Ebbers, F. G. Germain, G. Wichern, and J. Le Roux, “Keeping the balance: Anomaly score calculation for domain generalization,” inProc. ICASSP, 2025

    2025

  19. [27]

    Local density-based anomaly score normalization for do- main generalization,

    ——, “Local density-based anomaly score normalization for do- main generalization,”IEEE Trans. Audio, Speech, Lang. Process., vol. 33, 2025

    2025

  20. [28]

    Handling domain shifts for anomalous sound detection: A review of DCASE-related work,

    K. Wilkinghoff, T. Fujimura, K. Imoto, J. Le Roux, Z.-H. Tan, and T. Toda, “Handling domain shifts for anomalous sound detection: A review of DCASE-related work,” inProc. DCASE, 2025

    2025

  21. [29]

    ASDKit: A toolkit for comprehensive evaluation of anomalous sound detec- tion methods,

    T. Fujimura, K. Wilkinghoff, K. Imoto, and T. Toda, “ASDKit: A toolkit for comprehensive evaluation of anomalous sound detec- tion methods,” inProc. DCASE, 2025

    2025

  22. [30]

    Adjust- ing bias in anomaly scores via variance minimization for domain- generalized discriminative anomalous sound detection,

    M. Matsumoto, T. Fujimura, W. Huang, and T. Toda, “Adjust- ing bias in anomaly scores via variance minimization for domain- generalized discriminative anomalous sound detection,” inProc. DCASE, 2025

    2025

  23. [31]

    Description and discussion on DCASE2020 Challenge Task2: Unsupervised anomalous sound detection for machine condition monitoring,

    Y . Koizumiet al., “Description and discussion on DCASE2020 Challenge Task2: Unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2020

    2020

  24. [32]

    MIMII dataset: Sound dataset for malfunc- tioning industrial machine investigation and inspection,

    H. Purohitet al., “MIMII dataset: Sound dataset for malfunc- tioning industrial machine investigation and inspection,” inProc. DCASE, 2019

    2019

  25. [33]

    Toy- ADMOS: A dataset of miniature-machine operating sounds for anomalous sound detection,

    Y . Koizumi, S. Saito, H. Uematsu, N. Harada, and K. Imoto, “Toy- ADMOS: A dataset of miniature-machine operating sounds for anomalous sound detection,” inProc. WASPAA, 2019

    2019

  26. [34]

    MIMII DG: Sound dataset for malfunctioning industrial machine investigation and inspection for domain gener- alization task,

    K. Dohiet al., “MIMII DG: Sound dataset for malfunctioning industrial machine investigation and inspection for domain gener- alization task,” inProc. DCASE, 2022

    2022

  27. [35]

    ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, M. Yasuda, and S. Saito, “ToyADMOS2: Another dataset of miniature-machine operating sounds for anomalous sound detection under domain shift conditions,” inProc. DCASE, 2021

    2021

  28. [36]

    ToyADMOS2+: New Toyadmos data and benchmark results of the first-shot anomalous sound event detection baseline,

    N. Harada, D. Niizumi, D. Takeuchi, Y . Ohishi, and M. Yasuda, “ToyADMOS2+: New Toyadmos data and benchmark results of the first-shot anomalous sound event detection baseline,” inProc. DCASE, 2023

    2023

  29. [37]

    ToyADMOS2#: Yet another dataset for the DCASE2024 challenge task 2 first-shot anomalous sound detection,

    D. Niizumi, N. Harada, Y . Ohishi, D. Takeuchi, and M. Ya- suda, “ToyADMOS2#: Yet another dataset for the DCASE2024 challenge task 2 first-shot anomalous sound detection,” inProc. DCASE, 2024

    2024

  30. [38]

    IMAD-DS: A dataset for industrial multi-sensor anomaly detection under domain shift conditions,

    D. Albertini, F. Augusti, K. Esmer, A. Bernardini, and R. San- nino, “IMAD-DS: A dataset for industrial multi-sensor anomaly detection under domain shift conditions,” inProc. DCASE, 2024

    2024

  31. [39]

    ToyADMOS2025: The evaluation dataset for the DCASE2025T2 first-shot unsupervised anomalous sound detec- tion for machine condition monitoring,

    N. Harada, D. Niizumi, Y . Ohishi, D. Takeuchi, and M. Ya- suda, “ToyADMOS2025: The evaluation dataset for the DCASE2025T2 first-shot unsupervised anomalous sound detec- tion for machine condition monitoring,” inProc. DCASE, 2025

    2025

  32. [40]

    Analyzing a portion of the ROC curve,

    D. K. McClish, “Analyzing a portion of the ROC curve,”Medical decision making, vol. 9, no. 3, 1989

    1989

  33. [41]

    Self-supervised learning for anomalous sound detection,

    K. Wilkinghoff, “Self-supervised learning for anomalous sound detection,” inProc. ICASSP, 2024

    2024

  34. [42]

    AdaProj: Adaptively scaled angular margin subspace pro- jections for anomalous sound detection with auxiliary classifica- tion tasks,

    ——, “AdaProj: Adaptively scaled angular margin subspace pro- jections for anomalous sound detection with auxiliary classifica- tion tasks,” inProc. DCASE, 2024

    2024

  35. [43]

    Look, listen, and learn more: Design choices for deep audio embeddings,

    A. Cramer, H. Wu, J. Salamon, and J. P. Bello, “Look, listen, and learn more: Design choices for deep audio embeddings,” inProc. ICASSP, 2019

    2019

  36. [44]

    BEATs: Audio pre-training with acoustic tokeniz- ers,

    S. Chenet al., “BEATs: Audio pre-training with acoustic tokeniz- ers,” inProc. ICML, 2023

    2023

  37. [45]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmekeet al., “Audio set: An ontology and human-labeled dataset for audio events,” inProc. ICASSP, 2017

    2017

  38. [46]

    EAT: self- supervised pre-training with efficient audio transformer,

    W. Chen, Y . Liang, Z. Ma, Z. Zheng, and X. Chen, “EAT: self- supervised pre-training with efficient audio transformer,” inProc. IJCAI, 2024

    2024

  39. [47]

    Scaling up masked audio encoder learning for general audio clas- sification,

    H. Dinkel, Z. Yan, Y . Wang, J. Zhang, Y . Wang, and B. Wang, “Scaling up masked audio encoder learning for general audio clas- sification,” inProc. Interspeech, 2024

    2024

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.