Pith. sign in

REVIEW 3 major objections 6 minor 35 references

IMPACT: Industrial Machine Perception via Acoustic Cognitive Transformer

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A self-supervised transformer pretrained on 1,093 hours of industrial machine sound outperforms general-audio and respiratory-audio foundation models on 24 of 30 industrial classification and fault-detection tasks.

desk verdict The dataset is a real contribution; the model is EAT re-applied, and the 24/30 superiority claim is not yet supported by the numbers as printed. read the letter →

arxiv 2507.06481 v1 pith:NQZBHAT3 submitted 2025-07-09 cs.SD eess.AS

classification cs.SDeess.AS
keywords industrialmachinesoundacousticfoundationmodelself-supervisedlearningDINOSdatasetanomalousdetectiontransferstateclassificationmaskedspectrogrammodeling
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

The paper argues that industrial machine sounds are acoustically distinct from speech, music, and general environmental audio, and that this gap is why generic audio foundation models underperform in factories. To close it, the authors release DINOS, a corpus of 74,149 clips totaling over 1,093 hours of recordings from CNC cutting, laser powder bed fusion, directed energy deposition, cold spray, and shop-floor environments, captured with both stethoscope and microphone sensors. They also introduce IMPACT, a transformer pretrained on DINOS with a self-supervised objective that reconstructs masked spectrograms while aligning utterance-level representations. Benchmarking on 30 downstream tasks across four machine types, IMPACT is reported to achieve the highest F1 score on 24 of them, and the same-architecture ablation shows that fine-tuning AudioMAE on DINOS lifts ColdSpray F1 from 0.7598 to 0.8778. If right, the paper establishes that domain-specific industrial acoustic pretraining, not just larger general-audio data, is a path to scalable machine monitoring and fault detection.

What carries the argument

The load-bearing mechanism is the combination of a domain-matched pretraining corpus and a dual-objective self-supervised objective. IMPACT adopts the efficient audio transformer (EAT) student–teacher architecture: the student sees a log-Mel spectrogram with 70% of patches masked and must reconstruct it under a frame-level Huber loss while matching the teacher's average-pooled layer outputs on the unmasked input under an utterance-level loss weighted by $\lambda=0.1$. The teacher's weights are an exponential moving average of the student's, and gradients do not flow into the teacher. The paper's claim is that this captures both fine-grained spectral detail and global temporal structure, and that DINOS's mix of stethoscope and microphone recordings of repetitive machine kinematics is what makes the representations transfer to machine-state and fault-detection tasks.

What would settle it

Check whether the benchmark split is at clip or recording level: if one-second clips cut from the same continuous session appear in both train and test sets, re-run the benchmark keeping all clips from each session in the same split, and compare the F1 gaps. Independently, re-label the Yornew chatter states and cold-spray fault states from vibration sensors or process logs; if the acoustic labels do not match those references, the task definitions themselves are the confound rather than the model's representations.

Watch

Extended reading notes

Core claim

IMPACT is presented as the first foundation model trained specifically on industrial machine sound, and it is pretrained on DINOS, a dataset collected with both stethoscope and microphone sensors across CNC, directed energy deposition, laser powder bed fusion, cold spray, and shop-floor environments. The model adopts an efficient audio transformer student–teacher setup with a 70% masked-spectrogram input, jointly minimizing a frame-level Huber reconstruction loss and an utterance-level mean-squared-error alignment loss. On 30 downstream tasks—machine on/off classification, CNC operation mode recognition, 17-class machining state and chatter identification, and cold-spray fault detection across two sensor types—IMPACT attains the best mean F1 on 24 tasks. The paper attributes the transfer to the pretraining data: fine-tuning AudioMAE on the same DINOS subset raises its ColdSpray F1 from 0.7598 to 0.8778, and IMPACT still exceeds that level.

Load-bearing premise

The benchmark labels are accurate and the 20/80 splits do not leak adjacent one-second clips from the same continuous recording into both training and evaluation.

Editorial extensions

If this is right

  • Frozen IMPACT representations plus a single fine-tuned linear layer are enough to beat full linear-probe baselines, so downstream industrial tasks need only small labeled sets.
  • The fine-tuned AudioMAE results show that exposure to DINOS data, not the IMPACT architecture alone, drives most of the performance gain.
  • The benchmark establishes a reusable evaluation protocol and a new performance reference for 30 industrial sound tasks across four machines.
  • The dataset release gives the community a large-scale, open-access alternative to MIMII and ToyADMOS for pretraining and benchmarking industrial sound models.
  • General-purpose and respiratory-audio models plateau on fine-grained chatter and cold-spray faults, indicating that industrial monitoring requires its own foundation models.

Reading between the lines

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

  • If DINOS covers the diversity it claims, IMPACT-style pretraining could transfer to other rotating machinery such as pumps, fans, and gearboxes without per-machine training, a zero-shot or few-shot test the paper does not run.
  • The comparison against AudioMAE fine-tuned on DINOS isolates the data effect, but not the industrial-specific effect; pretraining an identical model on an equally sized random environmental corpus would sharpen the claim that industrial acoustics, not just in-domain scale, matter.
  • The Yornew chatter labels are acoustically inferred; ground-truthing them against vibration measurements or process logs would determine how much of the 17-task gain reflects true state separability rather than label artifacts.
  • Because IMPACT is an 18-million-parameter transformer, the same representations could plausibly be distilled for edge deployment, which the paper lists as future work.
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 / 6 minor

Summary. This paper presents DINOS, a new open-access dataset of 74,149 industrial audio clips (about 1,093 hours) collected from CNC, additive manufacturing, and cold-spray machines, and IMPACT, a transformer-based self-supervised model pretrained on a 121-hour subset of DINOS using an Efficient Audio Transformer (EAT) student-teacher objective with utterance- and frame-level losses. The authors benchmark IMPACT against seven baselines (OpenSMILE, CLAP, VGGish, OPERA-CT, OPERA-GT, AudioMAE pretrained, and AudioMAE fine-tuned on DINOS) using frozen-encoder linear probing with ten repeated 20/80 splits, reporting F1 and standard deviations on 30 downstream tasks across four machine types. They claim IMPACT outperforms existing models on 24 of 30 tasks and that domain-specific pretraining on DINOS is the key factor, supported by an AudioMAE fine-tuning ablation.

Significance. If the benchmark results hold, the paper makes two useful contributions: a relatively large open dataset of real industrial machine sounds with multiple sensor types, and a domain-specific audio foundation model that transfers to machine-state and fault-detection tasks better than general-audio and respiratory-audio models. The paper's methodology has notable strengths: frozen-encoder evaluation with ten repeated splits and reported standard deviations, benchmark machines disjoint from pretraining machines, and a same-architecture AudioMAE fine-tuning ablation that isolates the value of DINOS data. However, the central '24 of 30' superiority claim is not fully supported as written, and unresolved details about sample accounting and split construction could affect the rankings. The dataset and code release are valuable independent of the benchmark claim.

major comments (3)
  1. [5.1, Tables 1 and 2, and Section 5.3] The manuscript contains contradictory statements about the Yornew tasks. Table 2's header says T1–T22 are stethoscope-sourced, but Section 5.3 says the Yornew tasks use “multi-modal sounds recorded by both stethoscope and microphone sensors.” The sample counts for T6–T22 in Table 2 sum to 922, exactly twice the 461 stethoscope samples listed for VMC-300 in Table 1. Please clarify the sensor composition of each task and correct the header or the text accordingly. In addition, the 20/80 split is described only as random over “samples” (1-second clips), and the paper never states whether clips cut from the same continuous 59-second recording can appear in both train and test. Adjacent 1-second clips from the same machine state are near-duplicates; a clip-level split can inflate F1 for all models and can change the ordering between IMPACT and AudioMAE-FineT. A recording-level split, or an explicit justification for clip-level splitting, is needed to support the benchmark conclusions.
  2. [Abstract, Section 5.3, and Table 4] The headline claim that IMPACT “outperforms existing models on 24 tasks” is not supported by Table 4 as printed. IMPACT is exactly tied with CLAP and OPERA-CT on T1 and T2 (all 1.0000), and it is numerically below the best baseline on T9 (0.8419 vs CLAP 0.9239), T10 (0.8648 vs CLAP 0.8782), T11 (0.9475 vs AudioMAE-PreT. 0.9483), T15 (0.8458 vs CLAP 0.8564), T21 (0.9651 vs AudioMAE-FineT. 0.9665), and T22 (0.9465 vs CLAP 0.9536). At most 24 tasks show IMPACT as best or tied-best, and only 22 show strict superiority. Please correct the claim and add paired statistical tests across the ten splits (e.g., a paired t-test or Wilcoxon signed-rank test on the per-split F1 values) to justify “outperforms” rather than “matches or exceeds.”
  3. [5.1 and Table 2] The ground-truth labels for the Yornew chatter tasks (T12, T14, T16, T18–T22) and the ColdSpray fault states (depleted powder, powder clogging, no gas) are treated as definitive, but no verification protocol is described. Without a description of how chatter was confirmed (e.g., by vibration measurements or surface-finish inspection) and how the ColdSpray states were validated (e.g., by process logs or powder-flow sensors), the possibility of label noise that affects model rankings cannot be dismissed. Please document the labeling and verification procedure.
minor comments (6)
  1. [4.2 and 5.3] The pretraining objective is described as a “hybrid … contrastive and generative” strategy, but Section 4.2 defines only a frame-level Huber loss and an utterance-level MSE alignment loss; no contrastive loss term appears in L_total. Please correct the wording or specify the contrastive objective explicitly.
  2. [5.2] The paper reports F1 scores but does not specify whether the reported F1 is macro-averaged, micro-averaged, or weighted; please state the averaging scheme, especially for the multi-class Yornew tasks.
  3. [5.3, Yornew paragraph] The sentence “all pretrained models underperform (F1 < 0.86)” should read “all baseline pretrained models,” since IMPACT is also a pretrained model and achieves F1 > 0.89 on that machine.
  4. [Table 4] Please verify the bold formatting: as reproduced, IMPACT appears bolded on T1 and T2 where it ties at 1.0000, and on T21 where its F1 (0.9651) is below AudioMAE-FineT. (0.9665). If bold denotes “best or tied best,” state this explicitly in the table caption.
  5. [Global] There are typos: “fine-tunedd” appears in Sections 5.2 and 5.3, and “pre-train” appears in Section 5.3.
  6. [Abstract and Section 4.1] The abstract emphasizes the full 1,093-hour dataset, but pretraining uses only a 121-hour subset; please clarify in Section 4.1 that the remaining hours are reserved for benchmarking and future use.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: pretraining and benchmark data are disjoint, and the central transfer claim rests on external baselines and an independent same-architecture ablation, not on construction or self-citation.

full rationale

The paper's derivation chain is self-contained with respect to the claimed result. IMPACT is pretrained on DINOS using a self-supervised masked-autoencoding objective (L_total = L_f + lambda L_u), and the benchmark evaluation uses frozen encoders plus a linear probe trained on 20% of each downstream task and evaluated on the remaining 80%. None of these steps defines a target quantity in terms of the input it is supposed to predict, and no fitted parameter is renamed as a prediction. The benchmark machines (RenishawL, VF2, VMC-300, ColdSpray) are distinct from the pretraining machines (SK2540, RenishawR, FormAlloy, shop floor), and the paper explicitly states 'None of the benchmark samples are included in the pretraining data.' The cold-spray and CNC labels are process-set ground truths (RPM, MRR, chatter, normal/depleted/clogging/no-gas states) rather than model-derived labels. The same-architecture ablation in Section 5.3, in which AudioMAE is fine-tuned on DINOS and improves ColdSpray F1 from 0.7598 to 0.8778, provides independent evidence that the pretraining corpus, not the IMPACT architecture alone, drives transfer. The self-citations present ([20], [24]) support sensor selection and prior cold-spray diagnostic work; they are not load-bearing for the central claim that IMPACT outperforms other models, which is evaluated against externally released baselines (CLAP, VGGish, AudioMAE, OPERA-CT, OPERA-GT). One accuracy issue exists but is not circularity: the abstract's claim of '24 tasks' is not strictly supported by Table 4, since IMPACT ties on T1/T2 and is numerically lower than a baseline on T9, T10, T11, T15, T21, and T22. That is a correctness or reporting concern, not a circular-derivation concern. Overall, no step reduces by construction to its own inputs, so the circularity score is 0.

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

No new physical entities are introduced; the model name's 'Cognitive' is branding. The central claims rest on five unablated protocol hyperparameters and on undocumented assumptions about label accuracy, split granularity, and clip sufficiency. Self-citations [15, 20, 24, 34] provide prior-published support for the stethoscope sensing method, which counts as independent grounding for the data-collection approach, not circularity. The empirical performance claim additionally depends on the assumption that frozen-encoder linear probing is a faithful proxy for downstream usefulness.

free parameters (5)
  • loss weight lambda = 0.1 (Section 4.2)
    Weights the utterance-level loss against the frame-level reconstruction loss; selected by hand with no ablation, although the dual-objective design is claimed as the source of IMPACT's gains.
  • spectrogram masking ratio = 0.7 (Section 4.2)
    Fraction of patches masked for the student branch; inherited from EAT without re-tuning for industrial audio.
  • pretraining epochs = 10 (Appendix A.3)
    Training duration with no reported early stopping or convergence check.
  • pretraining subset size per category = 1,851 clips of 59 s each (Section 4.1)
    Only 7,404 of 74,149 clips (about 121 of 1,093 hours) are used for pretraining; the selection rule is unexplained.
  • linear probe training fraction = 20% per task, 10 runs (Section 5.1)
    Standard linear-probe protocol; for tasks with 24-78 samples this leaves 5-15 training samples and drives the large reported variances.
assumptions (5)
  • domain assumption EAT's student-teacher masked autoencoding is an effective base for industrial audio representation learning
    Adopted wholesale in Section 4.2; no EAT-on-general-audio control, so the architecture's contribution to the results is untested.
  • domain assumption Linear probing on frozen embeddings ranks representation quality
    Protocol of Section 5.2; rankings may differ under full fine-tuning.
  • domain assumption Yornew chatter and ColdSpray fault-state labels are accurate
    Section 5.1 and Table 2 define the states; no verification via vibration sensors, process logs, or expert annotation is described.
  • domain assumption Pretraining and benchmark clips are disjoint, and splits prevent leakage
    Section 5.1 asserts disjointness and a random 20/80 split, but clip construction from recordings and split granularity are not documented.
  • domain assumption One-second clips suffice for the target states
    All benchmark samples are 1-second clips (Table 1); transient fault signatures may be truncated, and no segment-length analysis is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IMPACT: Industrial Machine Perception via Acoustic Cognitive Transformer." pith.science (2026). https://pith.science/paper/NQZBHAT3

@misc{pith2026250706481,
  author       = {Pith},
  title        = {Pith review of: IMPACT: Industrial Machine Perception via Acoustic Cognitive Transformer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQZBHAT3}},
  note         = {Machine review of arXiv:2507.06481}
}
read the original abstract

Acoustic signals from industrial machines offer valuable insights for anomaly detection, predictive maintenance, and operational efficiency enhancement. However, existing task-specific, supervised learning methods often scale poorly and fail to generalize across diverse industrial scenarios, whose acoustic characteristics are distinct from general audio. Furthermore, the scarcity of accessible, large-scale datasets and pretrained models tailored for industrial audio impedes community-driven research and benchmarking. To address these challenges, we introduce DINOS (Diverse INdustrial Operation Sounds), a large-scale open-access dataset. DINOS comprises over 74,149 audio samples (exceeding 1,093 hours) collected from various industrial acoustic scenarios. We also present IMPACT (Industrial Machine Perception via Acoustic Cognitive Transformer), a novel foundation model for industrial machine sound analysis. IMPACT is pretrained on DINOS in a self-supervised manner. By jointly optimizing utterance and frame-level losses, it captures both global semantics and fine-grained temporal structures. This makes its representations suitable for efficient fine-tuning on various industrial downstream tasks with minimal labeled data. Comprehensive benchmarking across 30 distinct downstream tasks (spanning four machine types) demonstrates that IMPACT outperforms existing models on 24 tasks, establishing its superior effectiveness and robustness, while providing a new performance benchmark for future research.

Figures

Figures reproduced from arXiv: 2507.06481 by the authors.

Figure 1
Figure 1. System Overview. After curating data using stethoscope and microphone sensors on industrial machines, we pretrain IMPACT, a transformer-based industrial sound foundation model. Benchmarking evaluates eight sound foundation models on 30 industrial downstream tasks. provide a potential avenue to overcome these challenges. Leveraging large-scale self-supervised learning, foundation models can learn versatile representa… view at source ↗
Figure 2
Figure 2. Target Machines and Sound Sensor Placements. Locations of microphone (blue circle) and stethoscope (red circle) sensors on various industrial machines, (a) Yornew CNC machine (VMC￾300): A microphone is positioned beside the spindle, and a stethoscope sensor is attached to the rigid body beneath the CNC table. (b) Cold spray powder feeder (BaltiCold Spray LTD, CSM 108.2): A stethoscope sensor is attached to the feede… view at source ↗
Figure 3
Figure 3. Architecture of IMPACT. Overall pipeline with student-teacher branches.The student is trained using a reconstruction and alignment objective, while the teacher is updated via EMA. The detailed hyperparameters are presented in Appendix A.3. 5 Benchmarking 5.1 Benchmark Datasets [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Mean frequency spectra with standard deviation error bars for various machine operating states and sensor configurations. (a) Frequency response of the Renishaw AM machine during operational and idle states, measured using a stethoscope acoustic sensor. (b) Spectra of …
Figure 5
Figure 5. Figure 5: shows overall model performance per machine, while [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Graph of Per-Class Model Performance. Bar chart comparing the average F1 scores (± standard deviation) of all baseline models for each downstream task. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Confusion Matrices for RenishawL. Each cell in the table shows the count of instances predicted by the model for the corresponding class. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Confusion Matrices for VF2. Each cell in the table shows the count of instances predicted by the model for the corresponding class. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png]
Figure 9
Figure 9. Figure 9: Confusion Matrices for Yornew. Each cell in the table shows the count of instances predicted by the model for the corresponding class. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Confusion Matrices for ColdSpray. Each cell in the table shows the count of instances predicted by the model for the corresponding class. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    J. Antoni. Cyclic spectral analysis in practice.Mechanical Systems and Signal Processing, 21 (2):597–630, 2007

  2. [2]

    Baevski, Y

    A. Baevski, Y . Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations.Advances in neural information processing systems, 33: 12449–12460, 2020

  3. [3]

    D. A. Bies, C. H. Hansen, C. Q. Howard, and K. L. Hansen.Engineering noise control. CRC press, 2023

  4. [4]

    Brown, B

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

  5. [5]

    Chasmai, A

    M. Chasmai, A. Shepard, S. Maji, and G. Van Horn. The inaturalist sounds dataset.Advances in Neural Information Processing Systems, 37:132524–132544, 2024

  6. [6]

    W. Chen, Y . Liang, Z. Ma, Z. Zheng, and X. Chen. Eat: Self-supervised pre-training with efficient audio transformer.arXiv preprint arXiv:2401.03497, 2024

  7. [7]

    S. Chu, S. Narayanan, and C.-C. J. Kuo. Environmental sound recognition with time–frequency audio features.IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142– 1158, 2009

  8. [8]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo- gies, volume 1 (long and short papers), pages 4171–4186, 2019

Show all 35 references
  1. [9]

    K. Dohi, T. Nishida, H. Purohit, R. Tanabe, T. Endo, M. Yamamoto, Y . Nikaido, and Y . Kawaguchi. Mimii dg: Sound dataset for malfunctioning industrial machine investiga- tion and inspection for domain generalization task.arXiv preprint arXiv:2205.13879, 2022

  2. [10]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

  3. [11]

    Elizalde, S

    B. Elizalde, S. Deshmukh, M. Al Ismail, and H. Wang. Clap learning audio concepts from natural language supervision. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

  4. [12]

    Eyben, M

    F. Eyben, M. Wöllmer, and B. Schuller. Opensmile: the munich versatile and fast open-source audio feature extractor. InProceedings of the 18th ACM international conference on Multimedia, pages 1459–1462, 2010

  5. [13]

    J. F. Gemmeke, D. P. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter. Audio set: An ontology and human-labeled dataset for audio events. In2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776–780. I...

  6. [14]

    Gong, Y .-A

    Y . Gong, Y .-A. Chung, and J. Glass. Ast: Audio spectrogram transformer.arXiv preprint arXiv:2104.01778, 2021

  7. [15]

    C. Han, J. Lee, M. B. Jun, S. W. Lee, and H. Yun. Visual coating inspection framework via self-labeling and multi-stage deep learning strategies.Journal of Intelligent Manufacturing, pages 1–18, 2024

  8. [16]

    Heittola, A

    T. Heittola, A. Mesaros, A. Eronen, and T. Virtanen. Context-dependent sound event detection. EURASIP Journal on audio, speech, and music processing, 2013:1–13, 2013. 11

  9. [17]

    Hershey, S

    S. Hershey, S. Chaudhuri, D. P. Ellis, J. F. Gemmeke, A. Jansen, R. C. Moore, M. Plakal, D. Platt, R. A. Saurous, B. Seybold, et al. Cnn architectures for large-scale audio classification. In2017 ieee international conference on acoustics, speech and signal processing (icassp)...

  10. [18]

    W.-N. Hsu, B. Bolte, Y .-H. H. Tsai, K. Lakhotia, R. Salakhutdinov, and A. Mohamed. Hubert: Self-supervised speech representation learning by masked prediction of hidden units.IEEE/ACM transactions on audio, speech, and language processing, 29:3451–3460, 2021

  11. [19]

    Huang, H

    P.-Y . Huang, H. Xu, J. Li, A. Baevski, M. Auli, W. Galuba, F. Metze, and C. Feichtenhofer. Masked autoencoders that listen.Advances in Neural Information Processing Systems, 35: 28708–28720, 2022

  12. [20]

    E. Kim, J. Jeon, Y . Kim, H. Yun, J. Wellman, Y . W. Choi, S. W. Lee, M. B.-G. Jun, and J. Lee. Control-resilient roller wear prediction for thin wire flattening process via an internal sound-guided dynamic conditional network.International Journal of Precision Engineering and...

  13. [21]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment anything. InProceedings of the IEEE/CVF international conference on computer vision, pages 4015–4026, 2023

  14. [22]

    Koizumi, S

    Y . Koizumi, S. Saito, H. Uematsu, N. Harada, and K. Imoto. Toyadmos: A dataset of miniature- machine operating sounds for anomalous sound detection. In2019 IEEE Workshop on Ap- plications of Signal Processing to Audio and Acoustics (WASPAA), pages 313–317. IEEE, 2019

  15. [23]

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition.IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880–2894, 2020

  16. [24]

    J. Lee, S. Akin, Y . Sim, H. Lee, E. Kim, J. Nam, K. Song, and M. B. Jun. A stethoscope- guided interpretable deep learning framework for powder flow diagnosis in cold spray additive manufacturing.Manufacturing Letters, 41:1515–1525, 2024

  17. [25]

    Marchi, F

    E. Marchi, F. Vesperini, S. Squartini, and B. Schuller. Deep recurrent neural network-based autoencoders for acoustic novelty detection.Computational intelligence and neuroscience, 2017 (1):4694860, 2017

  18. [26]

    K. J. Piczak. Environmental sound classification with convolutional neural networks. In2015 IEEE 25th international workshop on machine learning for signal processing (MLSP), pages 1–6. IEEE, 2015

  19. [27]

    K. J. Piczak. Esc: Dataset for environmental sound classification. InProceedings of the 23rd ACM international conference on Multimedia, pages 1015–1018, 2015

  20. [28]

    Purohit, R

    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.arXiv preprint arXiv:1909.09347, 2019

  21. [29]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

  22. [30]

    R. B. Randall.Vibration-based condition monitoring: industrial, automotive and aerospace applications. John Wiley & Sons, 2021

  23. [31]

    Salamon, C

    J. Salamon, C. Jacoby, and J. P. Bello. A dataset and taxonomy for urban sound research. In Proceedings of the 22nd ACM international conference on Multimedia, pages 1041–1044, 2014

  24. [32]

    Schuller, S

    B. Schuller, S. Steidl, A. Batliner, J. Hirschberg, J. K. Burgoon, A. Baird, A. Elkins, Y . Zhang, E. Coutinho, and K. Evanini. The interspeech 2016 computational paralinguistics challenge: deception, sincerity and native language. 2016. 12

  25. [33]

    Sivasankaran and K

    S. Sivasankaran and K. Prabhu. Robust features for environmental sound classification. In2013 IEEE International Conference on Electronics, Computing and Communication Technologies, pages 1–6. IEEE, 2013

  26. [34]

    H. Yun, H. Kim, Y . H. Jeong, and M. B. Jun. Autoencoder-based anomaly detection of industrial robot arm using stethoscope based internal sound sensor.Journal of Intelligent Manufacturing, 34(3):1427–1444, 2023

  27. [35]

    on,” and all other segments as “off

    Y . Zhang, T. Xia, J. Han, Y . Wu, G. Rizos, Y . Liu, M. Mosuily, J. Ch, and C. Mascolo. Towards open respiratory acoustic foundation models: Pretraining and benchmarking.Advances in Neural Information Processing Systems, 37:27024–27055, 2024. 13 A Technical Appendices and Sup...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.