Pith. sign in

REVIEW 2 major objections 5 minor 63 references

ECHOv2: Two-Level Band-Splitting Representation Learning for Anomalous Sound Detection

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Teaching frequency bands to share context, not just encode themselves, improves machine anomalous sound detection across six years of DCASE benchmarks.

desk verdict Solid incremental ASD paper: explicit inter-band self-distillation on a band-splitting backbone plus a useful multi-year dual-protocol benchmark; gains are real but small and partly protocol-tied. read the letter →

arxiv 2607.10596 v1 pith:6QDIH3HW submitted 2026-07-12 eess.AS

Yucong Zhang , Juan Liu , Ming Li This is my paper
classification eess.AS
keywords anomaloussounddetectionband-splittingself-distillationaudiorepresentationlearningpre-trainedmodelsfrequencymodelingDCASEmachineconditionmonitoring
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

Machine anomalous sound detection needs representations that catch subtle spectral faults when only normal machine audio is available for training. General pre-trained audio models do not fully exploit how machine cues sit in different frequency regions. This paper extends a band-splitting backbone by keeping localized within-band learning and adding a second self-distillation stage: sub-band features exchange information through global context alignment, masked band reconstruction, and multiple summary tokens that aggregate low-, mid-, and high-frequency groups at controllable granularity. The authors also define a unified evaluation over DCASE 2020–2025 with two protocols—frozen embedding nearest-neighbor scoring and a light linear adaptation—so pre-trained backbones can be compared under consistent rules. Ablations and multi-year results show that explicit inter-band supervision and a moderate number of summary tokens improve overall scores over the prior band-splitting model and strong public audio baselines under both protocols.

What carries the argument

Two-level self-distillation on a band-splitting backbone: each sub-band is trained with masked reconstruction and context matching, then an inter-band branch aligns several frequency-region summary tokens and reconstructs masked band representations so cross-frequency dependencies become explicit during pre-training while inference still concatenates band features.

What would settle it

Re-run the same multi-year comparison of ECHOv2 versus its band-splitting predecessor with identical pre-training, but replace the benchmark’s nearest-neighbor cosine bank with a different standard ASD scorer (for example Gaussian density or reconstruction residual) under the same year-wise protocols; if the overall ranking and year-level gains vanish, the claim that the inter-band design improves representation quality is not supported.

Watch

Extended reading notes

Core claim

Structured cross-band modeling—localized intra-band self-distillation plus inter-band context alignment, masked sub-band reconstruction, and multi-summary-token aggregation—yields stronger anomalous-sound representations than band-splitting alone or general-purpose pre-trained audio models, as measured by both frozen-embedding discriminability and lightweight adaptation transfer on a unified DCASE 2020–2025 benchmark.

Load-bearing premise

That gains under one fixed nearest-neighbor cosine memory-bank scorer and one lightweight linear adapter, with each year’s official metadata matching and score aggregation, fairly isolate representation quality rather than how well each model fits that scoring recipe.

Editorial extensions

If this is right

  • ASD systems can improve representation quality with cross-band training-time supervision without changing the inference format of concatenated band features.
  • Three summary tokens strike a better balance for machine sounds than a single global token or six finer tokens under both evaluation protocols.
  • Inter-context alignment is the main driver of the inter-band gain; masked reconstruction and frequency positional embeddings act as secondary refinements.
  • A dual-protocol benchmark (frozen embeddings versus light adaptation) separates intrinsic discriminability from transferability when comparing audio backbones for ASD.
  • Open release of the model and the multi-year protocols makes future backbone comparisons on machine ASD more reproducible and less pipeline-dependent.

Reading between the lines

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

  • The same multi-summary-token cross-band idea may help other narrow-band anomaly tasks such as medical auscultation or industrial vibration monitoring.
  • Adaptive band partitions or data-dependent summary-token allocation could further help first-shot and domain-shift machine types the fixed design does not tailor.
  • If inter-band context mainly reorganizes the embedding geometry, denser or parametric scorers might enlarge or shrink the reported margins relative to nearest-neighbor cosine.
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

2 major / 5 minor

Summary. The paper proposes ECHOv2, extending the authors’ ECHO band-splitting backbone with a two-level self-distillation scheme: intra-band masked reconstruction and context alignment, plus an inter-band branch that performs global context alignment, masked sub-band reconstruction, and structured aggregation via multiple frequency-region summary tokens. The design aims to capture both fine-grained spectral patterns and cross-frequency dependencies for machine anomalous sound detection (ASD). The authors also introduce a unified ASD benchmark over DCASE 2020–2025 with two protocols—embedding-based nearest-neighbor cosine scoring of frozen representations and adaptation-based evaluation with a fixed linear projection—and report consistent but modest gains of ECHOv2 over ECHO-Small and strong public audio backbones (overall 62.63/64.52 vs 62.11/64.27), supported by ablations, summary-token sweeps, representation-type comparisons, and paired year-level t-tests. Models and benchmark code are open-sourced.

Significance. If the gains hold under broader scoring and adaptation choices, the work offers a clear, frequency-structured alternative to generic ViT-style audio pre-training for ASD and a reusable multi-year evaluation scaffold that the community currently lacks. Strengths include open release of model and benchmark, multi-year coverage under two complementary protocols, component ablations (Table V), summary-token granularity analysis (Table VI), and paired statistical tests (Table VII). The contribution is incremental over ECHO rather than transformative, and the absolute margins are small, but the combination of explicit inter-band supervision, structured aggregation, and a standardized ASD backbone benchmark is useful for the field.

major comments (2)
  1. §III-C/D and Fig. 6: The central claim that the two protocols isolate representation quality rests on a single fixed scorer (ℓ2-normalized nearest-neighbor cosine to a year-specific metadata-matched memory bank) and one fixed linear adapter (dim 256, 10 epochs, metadata-derived labels). Year-level margins are small; adaptation vs ECHO is not significant (Table VII, p=0.1843); removing inter-context returns embedding overall exactly to ECHO-Small (Table V). Without at least one alternative scorer (e.g., kNN with k>1, Mahalanobis, GMM, or a non-linear adapter), it remains possible that band-concatenation interacts favorably with this particular protocol rather than reflecting transferable discriminability. A short sensitivity experiment on 1–2 years would substantially strengthen the claim.
  2. Tables III–IV and VII: Gains over ECHO-Small are consistent but modest (≈0.5 overall points) and mixed year-wise (e.g., embedding 2023 slightly below ECHO). The paper should more carefully bound the claim—e.g., that inter-band supervision yields a small but statistically supported improvement under the stated protocols—rather than framing structured cross-band modeling as broadly “powerful” without quantifying practical impact or reporting confidence intervals / effect sizes alongside p-values.
minor comments (5)
  1. Table I and §II: Clarify whether the inter-band encoder and summary tokens are used only at training time or also at inference; the default downstream representation is band concatenation, but this is easy to miss.
  2. §II-E / Eq. (17): The two-stage schedule and equal λ splitting are stated, but a short justification or sensitivity note for the 400k+100k split and λ values would help reproducibility.
  3. §IV-A: Pre-training data scale is described as conservative relative to some baselines; a compact table of pre-training corpora and approximate hours for each compared model would make that comparison more transparent.
  4. Notation: FPE (Eq. 9) and summary-token grouping G_m (Eq. 16) are clear, but the relationship between stratified sampling (low/mid/high) and M summary tokens could be stated more explicitly in the text around Fig. 5.
  5. Minor polish: occasional spacing issues in URLs and table headers; ensure “DCASE 2020–2025” is consistent throughout.

Circularity Check

1 steps flagged · score 1.0 of 10

Empirical ASD representation paper; no derivation-by-construction. Mild self-citation of authors' ECHO backbone is normal extension, not load-bearing circularity.

  1. self citation load bearing [§I contributions; Table I; §II-A Recap of ECHO; Tables III–VI baselines]
    "We present ECHO as the band-splitting backbone of this journal work [43]. ECHO decomposes time–frequency inputs into sub-bands and learns frequency-localized representations through intra-band self-distillation. ... We extend ECHO to ECHOv2 to address its limitation in modeling cross-band dependencies."

    The paper's architecture and strongest baseline are the authors' prior ECHO model. This is ordinary self-extension, not a uniqueness theorem or forced prediction: gains of ECHOv2 over ECHO and over independent public backbones are measured on external DCASE data with ablations. Flagged only as mild self-citation dependence; it does not make the reported ASD scores true by construction.

full rationale

ECHOv2 is an empirical architecture-and-benchmark paper, not a first-principles derivation that redefines its target. The claimed chain is: (i) band-split spectrograms and train intra-band self-distillation (masked latent reconstruction + context alignment to EMA teacher targets); (ii) add an inter-band branch with context alignment, masked band reconstruction, FPE, and multi-summary-token aggregation; (iii) evaluate frozen and lightly adapted embeddings on external DCASE 2020–2025 data under fixed NN-cosine memory-bank and linear-adapter protocols against public baselines (BEATs, CED, Dasheng, EAT, FISHER). Training losses (Eqs. 5–6, 12–13, 15–17) optimize representation learning; evaluation metrics (AUC/pAUC under year-specific aggregation) are independent external scores, not algebraic rearrangements of those losses. Ablations (Tables V–VI) and paired t-tests (Table VII) are empirical sensitivity checks, not tautologies: removing inter-context returns embedding overall to ECHO-Small is evidence of contribution, not circular definition. Self-citation of ECHO [43] supplies the backbone and a strong baseline; that is ordinary incremental work and does not force the reported gains over public models or the multi-year ranking. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported to forbid alternatives, and no known empirical pattern is merely re-labeled. Protocol–model coupling (fixed scorer/adapter) is a validity concern for isolating representation quality, not circularity of the derivation chain. Score 1 reflects only the non-load-bearing ECHO self-citation; central claims rest on external benchmarks and open baselines.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central empirical claim rests on standard self-supervised audio modeling assumptions, DCASE evaluation conventions, and several design choices fixed by the authors (band width, summary-token count, loss weights, two-stage schedule, NN cosine scoring). No new physical entities are postulated; free parameters are architectural/training knobs selected for the reported runs rather than fitted constants in a closed-form law.

free parameters (4)
  • number of summary tokens M = 3 (default)
    Controls inter-band aggregation granularity; default M=3 chosen after comparing 1/3/6 and used for main claims.
  • sub-band width = 32
    Fixed frequency partition width for band splitting; determines the modeling units for both branches.
  • loss weights λ1–λ4 and two-stage schedule = 0.5/0.5 then 0.25 each; 400k+100k steps
    Stage-1 intra-only 400k steps with λ1=λ2=0.5; stage-2 joint 100k steps with equal 0.25 weights. These hand-set coefficients shape the learned representation.
  • inter-band encoder depth/heads and adaptation projection dim = 2 layers / 4 heads; proj dim 256
    Two Transformer layers, four heads; adaptation projection dimension 256. Architectural knobs that affect reported transfer scores.
assumptions (5)
  • domain assumption Teacher–student EMA self-distillation with masked reconstruction yields useful audio representations for downstream ASD.
    Core training paradigm inherited from ECHO/EAT-style SSL and used for both intra- and inter-band branches (§II-B, §II-C).
  • domain assumption Machine anomalies often appear as localized spectral changes, so frequency-band decomposition is an appropriate inductive bias.
    Motivation for band splitting and cross-band modeling in Introduction and methodology.
  • domain assumption Nearest-neighbor cosine distance in embedding space (with year-specific metadata matching) is a valid proxy for representation discriminability.
    Defines both evaluation protocols in §III-C/D and all main result tables.
  • domain assumption Official DCASE year-specific AUC/pAUC granularities and arithmetic/harmonic means are the right aggregation for multi-year comparison.
    Table II and §III-B fix the scoring rules used for overall claims.
  • standard math Transformer self-attention and sinusoidal frequency positional encodings are valid operators for band-token interaction.
    Used without derivation as standard sequence-model components in the inter-band encoder.
invented entities (2)
  • ECHOv2 two-level inter-band branch with multi-summary-token structured aggregation
    purpose: Explicitly model cross-frequency dependencies during pre-training while keeping band-concatenated features for inference.
    Architectural construct introduced by the paper; independent evidence is empirical ASD benchmark performance, not an external physical prediction.
  • Unified ASD backbone benchmark (embedding-based + adaptation-based protocols over DCASE 2020–2025) independent evidence
    purpose: Provide standardized comparison of frozen and lightly adapted pre-trained audio representations for ASD.
    Evaluation construct assembled from official DCASE settings; value is methodological standardization rather than a new natural entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECHOv2: Two-Level Band-Splitting Representation Learning for Anomalous Sound Detection." pith.science (2026). https://pith.science/paper/6QDIH3HW

@misc{pith2026260710596,
  author       = {Pith},
  title        = {Pith review of: ECHOv2: Two-Level Band-Splitting Representation Learning for Anomalous Sound Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6QDIH3HW}},
  note         = {Machine review of arXiv:2607.10596}
}
read the original abstract

Machine anomalous sound detection (ASD) requires robust audio representations capable of capturing subtle deviations in machine sounds under limited supervision. Existing pre-trained audio backbones do not fully capture frequency-specific characteristics of machine sounds. To address this, we propose ECHOv2, a band-splitting model that learns localized intra-band representations to capture fine-grained spectral patterns while also incorporating a two-level self-distillation strategy with explicit inter-band supervision to model cross-frequency dependencies. The inter-band branch performs global context alignment and masked sub-band reconstruction, and multiple summary tokens are introduced for structured aggregation with controllable frequency granularity, enabling region-aware interaction across sub-bands during training. This design allows ECHOv2 to robustly handle diverse machine types and noisy operating conditions while maintaining stable representation quality. To enable fair and consistent evaluation of pre-trained audio backbones, we establish a unified ASD benchmark over DCASE 2020-2025 with two complementary protocols: embedding-based evaluation for frozen representation discriminability and adaptation-based evaluation for downstream transferability. Ablation studies confirm the effectiveness of intra-band learning, inter-band supervision, and structured aggregation granularity for robust ASD representation learning. These findings demonstrate that structured cross-band modeling provides a powerful and adaptable framework for ASD representation learning and can serve as a strong foundation for future research. The model and benchmark are fully open-sourced at https://github.com/yucongzh/ECHOv2 and https://github.com/yucongzh/ASD_Benchmark to promote reproducible research.

Figures

Figures reproduced from arXiv: 2607.10596 by the authors.

Figure 1
Figure 1. Inference process of the ECHO band-splitting backbone. The input spectrogram is divided into frequency sub-bands, which are encoded by a shared [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Intra-band self-distillation for a single sub-band. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Inter-band self-distillation over the sequence of sub-band representa [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Structured inter-band learning with multiple summary tokens. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Band-level, student-side view of the proposed training pipeline. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Embedding-based anomaly scoring protocol. Normal training samples are first encoded into reference embeddings to construct a reference bank. For [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 1 linked inside Pith

  1. [1]

    Description and discussion on dcase2020 challenge task2: Unsupervised anomalous sound detection for machine condition monitoring,

    Y . Koizumi, Y . Kawaguchi, K. Imoto, T. Nakamura, Y . Nikaido, R. Tanabeet al., “Description and discussion on dcase2020 challenge task2: Unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2020, pp. 81–85

    2020

  2. [2]

    Description and discussion on dcase 2021 challenge task 2: Unsupervised anomalous detection for machine condition monitoring under domain shifted conditions,

    Y . Kawaguchi, K. Imoto, Y . Koizumi, N. Harada, D. Niizumi, K. Dohi et al., “Description and discussion on dcase 2021 challenge task 2: Unsupervised anomalous detection for machine condition monitoring under domain shifted conditions,” inProc. DCASE, 2021, pp. 186–190

    2021

  3. [3]

    Description and discussion on dcase 2022 challenge task 2: Un- supervised anomalous sound detection for machine condition monitoring applying domain generalization techniques,

    K. Dohi, K. Imoto, N. Harada, D. Niizumi, Y . Koizumi, T. Nishida et al., “Description and discussion on dcase 2022 challenge task 2: Un- supervised anomalous sound detection for machine condition monitoring applying domain generalization techniques,” inProc. DCASE, 2022

    2022

  4. [4]

    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, pp. 31–35

    2023

  5. [5]

    Description and discussion on dcase 2024 challenge task 2: First- shot unsupervised anomalous sound detection for machine condition monitoring,

    T. Nishida, N. Harada, D. Niizumi, D. Albertini, R. Sannino, S. Pradolini et al., “Description and discussion on dcase 2024 challenge task 2: First- shot unsupervised anomalous sound detection for machine condition monitoring,” inProc. DCASE, 2024, pp. 111–115

    2024

  6. [6]

    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

  7. [7]

    Anomalous sound detection based on interpolation deep neural network,

    K. Suefusa, T. Nishida, H. Purohit, R. Tanabe, T. Endo, and Y . Kawaguchi, “Anomalous sound detection based on interpolation deep neural network,” inProc. ICASSP, 2020, pp. 271–275

    2020

  8. [8]

    Anomaly detection in raw audio using deep autoregressive networks,

    E. Rushe and B. Mac Namee, “Anomaly detection in raw audio using deep autoregressive networks,” inProc. ICASSP, 2019, pp. 3597–3601

    2019

Show all 63 references
  1. [9]

    Unsupervised anomaly detection and localization of machine audio: A gan-based approach,

    A. Jiang, W.-Q. Zhang, Y . Deng, P. Fan, and J. Liu, “Unsupervised anomaly detection and localization of machine audio: A gan-based approach,” inProc. ICASSP, 2023, pp. 1–5

    2023

  2. [10]

    Unsupervised anomalous sound detection using self- supervised classification and group masked autoencoder for density estimation,

    R. Giri, S. V . Tenneti, K. Helwani, F. Cheng, U. Isik, and A. Kr- ishnaswamy, “Unsupervised anomalous sound detection using self- supervised classification and group masked autoencoder for density estimation,”Challenge on Detection and Classification of Acoustic Scenes and Ev...

    2020

  3. [11]

    Self- supervised representation learning for unsupervised anomalous sound detection under domain shift,

    H. Chen, Y . Song, L.-R. Dai, I. McLoughlin, and L. Liu, “Self- supervised representation learning for unsupervised anomalous sound detection under domain shift,” inProc. ICASSP, 2022, pp. 471–475

    2022

  4. [12]

    Outlier-aware Inlier Modeling and Multi-scale Scoring for Anomalous Sound Detection via Multitask Learning,

    Y . Zhang, S. Hongbin, Y . Wan, and M. Li, “Outlier-aware Inlier Modeling and Multi-scale Scoring for Anomalous Sound Detection via Multitask Learning,” inInterspeech 2023, 2023, pp. 5381–5385

    2023

  5. [13]

    Self-supervised learning for anomalous sound detec- tion,

    K. Wilkinghoff, “Self-supervised learning for anomalous sound detec- tion,” inProc. ICASSP, 2024, pp. 276–280

    2024

  6. [14]

    Adaptive prototype learning for anomalous sound detection with partially known attributes,

    A. Jiang, X. Zheng, B. Han, Y . Qiu, P. Fan, W.-Q. Zhang, C. Lu, and J. Liu, “Adaptive prototype learning for anomalous sound detection with partially known attributes,” inProc. ICASSP, 2025, pp. 1–5

    2025

  7. [15]

    Self-supervised acoustic anomaly detec- tion via contrastive learning,

    H. Hojjati and N. Armanfard, “Self-supervised acoustic anomaly detec- tion via contrastive learning,” inProc. ICASSP, 2022, pp. 3253–3257

    2022

  8. [16]

    Anomalous sound detection using audio representation with machine id based contrastive learning pretraining,

    J. Guan, F. Xiao, Y . Liu, Q. Zhu, and W. Wang, “Anomalous sound detection using audio representation with machine id based contrastive learning pretraining,” inProc. ICASSP, 2023, pp. 1–5

    2023

  9. [17]

    Joint generative-contrastive representation learning for anomalous sound detection,

    X.-M. Zeng, Y . Song, Z. Zhuo, Y . Zhou, Y .-H. Li, H. Xue, L.-R. Dai, and I. McLoughlin, “Joint generative-contrastive representation learning for anomalous sound detection,” inProc. ICASSP, 2023, pp. 1–5

    2023

  10. [18]

    Robust Prototype Learning for Anomalous Sound Detection,

    X.-M. Zeng, Y . Song, I. McLoughlin, L. Liu, and L.-R. Dai, “Robust Prototype Learning for Anomalous Sound Detection,” inInterspeech 2023, 2023, pp. 261–265

    2023

  11. [19]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthineret al., “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. ICLR, 2021

    2021

  12. [20]

    AST: Audio Spectrogram Trans- former,

    Y . Gong, Y .-A. Chung, and J. Glass, “AST: Audio Spectrogram Trans- former,” inInterspeech 2021, 2021, pp. 571–575

    2021

  13. [21]

    Masked autoencoders that listen,

    P.-Y . Huang, H. Xu, J. Li, A. Baevski, M. Auli, W. Galubaet al., “Masked autoencoders that listen,” inProc. NeurIPS, vol. 35, 2022, pp. 28 708–28 720

    2022

  14. [22]

    BEATs: Audio pre-training with acoustic tokenizers,

    S. Chen, Y . Wu, C. Wang, S. Liu, D. Tompkins, Z. Chenet al., “BEATs: Audio pre-training with acoustic tokenizers,” inProc. ICML, 2023, pp. 5178–5193

    2023

  15. [23]

    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, pp. 3807–3815

    2024

  16. [24]

    CED: Consistent ensemble distillation for audio tagging,

    H. Dinkel, Y . Wang, Z. Yan, J. Zhang, and Y . Wang, “CED: Consistent ensemble distillation for audio tagging,” inProc. ICASSP, 2024, pp. 291–295

    2024

  17. [25]

    Scaling up masked audio encoder learning for general audio classification,

    H. Dinkel, Z. Yan, Y . Wang, J. Zhang, Y . Wang, and B. Wang, “Scaling up masked audio encoder learning for general audio classification,” in Proc. Interspeech, 2024, pp. 547–551

    2024

  18. [26]

    Exploring large scale pre-trained models for robust machine anomalous sound detection,

    B. Han, Z. Lv, A. Jiang, W. Huang, Z. Chen, Y . Denget al., “Exploring large scale pre-trained models for robust machine anomalous sound detection,” inProc. ICASSP, 2024, pp. 1326–1330

    2024

  19. [27]

    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. ASLP, 2025

    2025

  20. [28]

    Anopatch: Towards better consistency in machine anomalous sound detection,

    A. Jiang, B. Han, Z. Lv, Y . Deng, W.-Q. Zhang, X. Chenet al., “Anopatch: Towards better consistency in machine anomalous sound detection,” inProc. Interspeech, 2024, pp. 107–111

    2024

  21. [29]

    Lora: Low-rank adaptation of large language models

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chenet al., “Lora: Low-rank adaptation of large language models.” 2022

    2022

  22. [30]

    Improving anomalous sound detection via low-rank adaptation fine- tuning of pre-trained audio models,

    X. Zheng, A. Jiang, B. Han, Y . Qian, P. Fan, J. Liu, and W.-Q. Zhang, “Improving anomalous sound detection via low-rank adaptation fine- tuning of pre-trained audio models,” inProc. SLT, 2024, pp. 969–974

    2024

  23. [31]

    Anomalous Sound Detection Using Self-Attention-Based Frequency Pattern Analysis of Machine Sounds,

    H. Zhang, J. Guan, Q. Zhu, F. Xiao, and Y . Liu, “Anomalous Sound Detection Using Self-Attention-Based Frequency Pattern Analysis of Machine Sounds,” inProc. Interspeech, 2023, pp. 336–340

    2023

  24. [32]

    Robust anomaly sound detection framework for machine condition monitoring,

    Y . Zeng, H. Liu, L. Xu, Y . Zhou, and L. Gan, “Robust anomaly sound detection framework for machine condition monitoring,” inProc. DCASE, 2022

    2022

  25. [33]

    Explaining the decisions of anomalous sound detectors,

    K. T. Mai, T. Davies, L. Griffi, E. Benetoset al., “Explaining the decisions of anomalous sound detectors,” inProc. DCASE

  26. [34]

    A dual-path framework with frequency-and-time excited network for anomalous sound detec- tion,

    Y . Zhang, J. Liu, Y . Tian, H. Liu, and M. Li, “A dual-path framework with frequency-and-time excited network for anomalous sound detec- tion,” inProc. ICASSP, 2024, pp. 1266–1270

    2024

  27. [35]

    A lightweight framework for unsupervised anomalous sound detection based on selective learning of time-frequency domain features,

    Y . Wang, Q. Zhang, W. Zhang, and Y . Zhang, “A lightweight framework for unsupervised anomalous sound detection based on selective learning of time-frequency domain features,”Applied Acoustics, vol. 228, p. 110308, 2025

    2025

  28. [36]

    Enhancing reliability through interpretability: A comprehensive survey of interpretable intelligent fault diagnosis in rotating machinery,

    G. Chen, J. Yuan, Y . Zhang, H. Zhu, R. Huang, F. Wang, and W. Li, “Enhancing reliability through interpretability: A comprehensive survey of interpretable intelligent fault diagnosis in rotating machinery,”IEEE access, vol. 12, pp. 103 348–103 379, 2024

    2024

  29. [37]

    Explainable artificial intelligence based intelligent fault diagnosis: A systematic review from applications to insights,

    T. Li, J. Chen, T. Liu, C. Sun, Z. Zhao, X. Chen, and R. Yan, “Explainable artificial intelligence based intelligent fault diagnosis: A systematic review from applications to insights,”Reliability Engineering & System Safety, p. 111935, 2025

    2025

  30. [38]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez et al., “Attention is all you need,”Proc. NeurIPS, vol. 30, 2017

    2017

  31. [39]

    Anomalous sound detection using spectral-temporal information fusion,

    Y . Liu, J. Guan, Q. Zhu, and W. Wang, “Anomalous sound detection using spectral-temporal information fusion,” inProc. ICASSP, 2022, pp. 816–820

    2022

  32. [40]

    Multi-scale scanning network for machine anomalous sound detection,

    Y . Zhang, J. Liu, and M. Li, “Multi-scale scanning network for machine anomalous sound detection,” inProc. ICONIP, 2025, pp. 303–316

    2025

  33. [41]

    Multi-spectral and multi-temporal features fusion with se network for anomalous sound detection,

    D. Kong, H. Yu, and G. Yuan, “Multi-spectral and multi-temporal features fusion with se network for anomalous sound detection,”IEEE Access, vol. 12, pp. 167 262–167 277, 2024

    2024

  34. [42]

    Estm: An enhanced dual-branch spectral-temporal mamba for anomalous sound detection,

    C. Ma, P. Jia, H. Guo, and W. Yang, “Estm: An enhanced dual-branch spectral-temporal mamba for anomalous sound detection,”IEEE SPL, 2025

    2025

  35. [43]

    Echo: Frequency-aware hierarchical encoding for variable-length signals,

    Y . Zhang, J. Liu, and M. Li, “Echo: Frequency-aware hierarchical encoding for variable-length signals,” inProc. ICASSP, 2026

    2026

  36. [44]

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

    P. Fan, A. Jiang, S. Zhang, Z. Lv, B. Han, X. Zhenget al., “FISHER: A foundation model for multi-modal industrial signal comprehensive representation,”arXiv preprint arXiv:2507.16696, 2025

    2025 arXiv

  37. [45]

    Hear: Holistic evaluation of audio representations,

    J. Turian, J. Shier, H. R. Khan, B. Raj, B. W. Schuller, C. J. Steinmetz, C. Malloy, G. Tzanetakis, G. Velarde, K. McNallyet al., “Hear: Holistic evaluation of audio representations,” inProc. NeurIPS, 2022, pp. 125– 145

    2022

  38. [46]

    SUPERB: Speech Processing Universal PERformance Benchmark,

    S. wen Yang, P.-H. Chi, Y .-S. Chuang, C.-I. J. Lai, K. Lakhotia, Y . Y . Lin, A. T. Liu, J. Shi, X. Chang, G.-T. Lin, T.-H. Huang, W.-C. Tseng, K. tik Lee, D.-R. Liu, Z. Huang, S. Dong, S.-W. Li, S. Watanabe, A. Mohamed, and H. yi Lee, “SUPERB: Speech Processing Universal PER...

    2021

  39. [47]

    X-ares: A comprehensive framework for assessing audio encoder performance,

    J. Zhang, H. Dinkel, Y . Niu, C. Liu, S. Cheng, A. Zhao, and J. Luan, “X-ares: A comprehensive framework for assessing audio encoder performance,” inProc. Interspeech, 2025

    2025

  40. [48]

    Mimii dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,

    H. Purohit, R. Tanabe, K. Ichige, T. Endo, Y . Nikaido, K. Suefusa, and Y . Kawaguchi, “Mimii dataset: Sound dataset for malfunctioning industrial machine investigation and inspection,” inProc. DCASE, 2019, pp. 209–213

    2019

  41. [49]

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

    Y . Koizumi, S. Saito, H. Uematsu, N. Harada, and K. Imoto, “Toyadmos: A dataset of miniature-machine operating sounds for anomalous sound detection,” inProc. WASPAA, 2019, pp. 313–317

    2019

  42. [50]

    Mimii due: Sound dataset for malfunctioning industrial machine investigation and inspection with domain shifts due to changes in operational and environmental conditions,

    R. Tanabe, H. Purohit, K. Dohi, T. Endo, Y . Nikaido, T. Nakamura, and Y . Kawaguchi, “Mimii due: Sound dataset for malfunctioning industrial machine investigation and inspection with domain shifts due to changes in operational and environmental conditions,” inProc. WASPAA, 20...

    2021

  43. [51]

    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

  44. [52]

    Mimii dg: Sound dataset for mal- functioning industrial machine investigation and inspection for domain generalization task,

    K. Dohi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, Y . Nikaido, and Y . Kawaguchi, “Mimii dg: Sound dataset for mal- functioning industrial machine investigation and inspection for domain generalization task,” inProc. DCASE, 2022, pp. 31–35

    2022

  45. [53]

    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, pp. 41–45

    2023

  46. [54]

    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. Yasuda, “Toyadmos2#: Yet another dataset for the dcase2024 challenge task 2 first-shot anomalous sound detection,” inProc. DCASE, 2024, pp. 106– 110

    2024

  47. [55]

    Toyadmos2025: The evaluation dataset for the dcase2025t2 first-shot unsupervised anomalous sound detection for machine condition moni- toring,

    N. Harada, D. Niizumi, Y . Ohishi, D. Takeuchi, and M. Yasuda, “Toyadmos2025: The evaluation dataset for the dcase2025t2 first-shot unsupervised anomalous sound detection for machine condition moni- toring,” inProc. DCASE, 2025, pp. 230–234

    2025

  48. [56]

    Imad-ds: A dataset for industrial multi-sensor anomaly detection under domain shift conditions,

    D. Albertini, F. Augusti, K. Esmer, A. Bernardini, and R. Sannino, “Imad-ds: A dataset for industrial multi-sensor anomaly detection under domain shift conditions,” inProc. DCASE, 2024, pp. 1–5

    2024

  49. [57]

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

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Mooreet al., “Audio set: An ontology and human-labeled dataset for audio events,” inProc. ICASSP, 2017, pp. 776–780

    2017

  50. [58]

    Vggsound: A large- scale audio-visual dataset,

    H. Chen, W. Xie, A. Vedaldi, and A. Zisserman, “Vggsound: A large- scale audio-visual dataset,” inProc. ICASSP, 2020, pp. 721–725

    2020

  51. [59]

    Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning,

    S. Lee, J. Chung, Y . Yu, G. Kim, T. Breuel, G. Chechik, and Y . Song, “Acav100m: Automatic curation of large-scale datasets for audio-visual video representation learning,” inProc. ICCV, 2021, pp. 10 274–10 284. JOURNAL OF LATEX CLASS FILES, VOL. 18, NO. 9, SEPTEMBER 2020 13

    2021

  52. [60]

    The mtg- jamendo dataset for automatic music tagging,

    D. Bogdanov, M. Won, P. Tovstogan, A. Porter, and X. Serra, “The mtg- jamendo dataset for automatic music tagging,” inProc. ICML, 2019

    2019

  53. [61]

    Music4all: A new music database and its applications,

    I. A. P. Santana, F. Pinhelli, J. Donini, L. Catharin, R. B. Mangolin, V . D. Feltrimet al., “Music4all: A new music database and its applications,” inProc. IWSSIP, 2020, pp. 399–404

    2020

  54. [62]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inProc. ICLR, 2019

    2019

  55. [63]

    Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio- language multimodal research,

    X. Mei, C. Meng, H. Liu, Q. Kong, T. Ko, C. Zhaoet al., “Wavcaps: A chatgpt-assisted weakly-labelled audio captioning dataset for audio- language multimodal research,”IEEE/ACM Trans. ASLP, vol. 32, pp. 3339–3354, 2024. Yucong Zhang(Graduate Student Member, IEEE) received the B...

    2024

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.