pith. machine review for the scientific record. sign in

arxiv: 2605.06309 · v2 · submitted 2026-05-07 · 💻 cs.CL

Recognition: unknown

MultiLinguahah : A New Unsupervised Multilingual Acoustic Laughter Segmentation Method

Callejas Sofia , Gomez Nahuel , Pelachaud Catherine , Ravenet Brian , Barriere Valentin
Authors on Pith no claims yet

Pith reviewed 2026-05-14 21:14 UTC · model grok-4.3

classification 💻 cs.CL
keywords laughter detectionunsupervised segmentationmultilingual audioanomaly detectionIsolation ForestBYOL-A representationsnon-English performance
0
0 comments X

The pith

An unsupervised anomaly detection method segments laughter in audio across languages using BYOL-A representations and Isolation Forest.

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

The paper frames laughter segmentation as an unsupervised anomaly detection task on energy-segmented audio sequences. It applies an Isolation Forest to features extracted by a BYOL-A encoder pretrained on general audio, avoiding any labeled examples or language-specific training. This setup is tested against existing laughter detection algorithms on four datasets covering stand-up comedy, sitcoms, and general short clips from AudioSet. The results indicate that prior methods, tuned mainly on English data, underperform in non-English settings while this approach handles multilingual cases more effectively. A reader would care because laughter is a universal social signal yet current tools remain limited by their reliance on costly annotations and English-centric training.

Core claim

The central claim is that laughter can be segmented unsupervised across languages by treating energy-based audio segments as anomalies and classifying them with an Isolation Forest applied to representations from a BYOL-A encoder. This yields better performance than state-of-the-art laughter detection methods on non-English portions of stand-up comedy, sitcom, and AudioSet data without requiring manual labels or language tuning.

What carries the argument

Isolation Forest classifier applied to BYOL-A learned representations of energy-segmented audio sequences, with laughter treated as the anomalous class.

If this is right

  • State-of-the-art methods remain limited outside English because they depend on language-specific labeled training.
  • The method requires no manual annotation, enabling application to new languages and audio sources.
  • It maintains accuracy on diverse inputs such as comedy performances and short general audio clips.
  • Anomaly detection on pretrained audio features can substitute for supervised classification in this domain.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same framing could apply to other universal non-verbal vocalizations such as sighs or gasps.
  • Real-time deployment on multilingual video platforms would become feasible without per-language retraining.
  • Extending evaluation to additional low-resource languages would test how far general audio pretraining generalizes.

Load-bearing premise

Representations learned by BYOL-A on general audio will reliably mark laughter as anomalies across languages and recording conditions without any language-specific tuning.

What would settle it

Performance on a held-out non-English dataset with varied noise levels and recording conditions drops below the baselines, showing the anomaly separation does not hold.

Figures

Figures reproduced from arXiv: 2605.06309 by Barriere Valentin, Callejas Sofia, Gomez Nahuel, Pelachaud Catherine, Ravenet Brian.

Figure 1
Figure 1. Figure 1: We first remove the voices from the laughter through channel subtraction or audio source separation (§2.1), second then we segment the audio into events using an energy-based threshold (§2.2), third we encode the audio using a pre-trained model (§2.3), and finally we detect laughter using an anomaly detection algorithm based on Isolation Forest (§2.4). lar, we used a basic deep learning model based on a de… view at source ↗
Figure 2
Figure 2. Figure 2: shows the F1 score of MultiLinguahah vs Omine et al.’s method on the Standup4AI dataset, with respect to the laughter duration. It is possible to see that the proposed method outperforms the baseline, in particular when the laughter events last longer. Once again, we believe that the pretrained ASR backbone used by Omine et al. works against it. Indeed, as soon as the laughter is too long or the speech is … view at source ↗
read the original abstract

Laughter is a social non-vocalization that is universal across cultures and languages, and is crucial for human communication, including social bonding and communication signaling. However, detecting laughter in audio is a challenging task, and segmenting is even more difficult. Currently, Machine Learning methods generally rely on costly manual annotation, and their datasets are mostly based on English contexts. Thus, we propose an unsupervised multilingual method that sets up the laughter segmentation task as an anomaly detection of energy-based segmented audio sequences. Our method applies an Isolation Forest on audio representations learned from BYOL-A encoder. We compare our method with several state-of-the-art laughter detection algorithms on four datasets, including stand-up comedy, sitcoms, and general short audio from AudioSet. Our results show that state-of-the-art methods are not optimized for multilingual contexts, while our method outperforms them in non-English settings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes MultiLinguahah, an unsupervised multilingual method for acoustic laughter segmentation. It frames the task as anomaly detection: energy-based audio segments are encoded with a frozen BYOL-A model and scored via Isolation Forest, with laughter treated as the anomalous class. The approach is evaluated against several supervised and unsupervised baselines on four datasets (stand-up comedy, sitcoms, and AudioSet subsets), with the central claim being superior performance over existing methods in non-English settings.

Significance. If the empirical claims hold under scrutiny, the work would offer a practical advance by removing the need for language-specific labeled data in laughter detection, a task relevant to social signal processing and conversational AI. The use of self-supervised BYOL-A representations is a methodological strength that could support cross-lingual generalization, though this remains to be demonstrated beyond the reported datasets.

major comments (3)
  1. [§3] §3 (Method): The core assumption that laughter occupies a reliably outlying region in BYOL-A space is load-bearing for the anomaly-detection framing, yet the manuscript provides no score histograms, density-conditioned precision-recall curves, or ablation on laughter frequency. In high-density stand-up comedy and sitcom data this assumption is particularly fragile and must be directly tested.
  2. [§4] §4 (Experiments): The reported gains in non-English conditions are presented without error bars, statistical significance tests, or per-dataset laughter-density statistics. Without these, it is impossible to determine whether the outperformance is robust or driven by post-hoc energy thresholding or dataset-specific artifacts.
  3. [§3.2] §3.2 (BYOL-A usage): The description does not state whether the BYOL-A encoder is used strictly frozen, how frame-level embeddings are aggregated over variable-length energy segments, or which layer is extracted. These choices directly affect the multilingual claim and must be specified for reproducibility.
minor comments (2)
  1. [Title] The title contains an apparent typographical error ('MultiLinguahah'); this should be corrected for clarity.
  2. [Tables] Table captions and axis labels in the results section should explicitly indicate language labels and laughter density per dataset to aid interpretation.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. These points help strengthen the methodological clarity and empirical rigor of the manuscript. We address each major comment below and will revise the paper to incorporate the suggested improvements.

read point-by-point responses

  1. Referee: [§3] §3 (Method): The core assumption that laughter occupies a reliably outlying region in BYOL-A space is load-bearing for the anomaly-detection framing, yet the manuscript provides no score histograms, density-conditioned precision-recall curves, or ablation on laughter frequency. In high-density stand-up comedy and sitcom data this assumption is particularly fragile and must be directly tested.

    Authors: We agree that direct validation of the anomaly assumption is essential, particularly for high-density datasets. In the revised manuscript we will add score histograms comparing laughter versus non-laughter segments in BYOL-A space, density-conditioned precision-recall curves, and an ablation that varies laughter frequency (by subsampling) to demonstrate that Isolation Forest continues to separate the classes reliably. revision: yes

  2. Referee: [§4] §4 (Experiments): The reported gains in non-English conditions are presented without error bars, statistical significance tests, or per-dataset laughter-density statistics. Without these, it is impossible to determine whether the outperformance is robust or driven by post-hoc energy thresholding or dataset-specific artifacts.

    Authors: We accept that the current experimental presentation lacks sufficient statistical support. We will augment §4 with error bars (standard deviation across runs or folds), appropriate statistical significance tests (e.g., paired Wilcoxon signed-rank tests) between MultiLinguahah and all baselines, and explicit per-dataset laughter-density statistics. These additions will allow readers to assess robustness independently of energy thresholding choices. revision: yes

  3. Referee: [§3.2] §3.2 (BYOL-A usage): The description does not state whether the BYOL-A encoder is used strictly frozen, how frame-level embeddings are aggregated over variable-length energy segments, or which layer is extracted. These choices directly affect the multilingual claim and must be specified for reproducibility.

    Authors: We will expand §3.2 with the missing implementation details. The BYOL-A encoder is used strictly frozen; frame-level embeddings are aggregated via mean pooling across the variable-length energy segments; and we extract the final-layer representations. These clarifications will be inserted verbatim to guarantee reproducibility of the multilingual results. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the unsupervised anomaly-detection pipeline

full rationale

The paper presents a direct application of pre-trained BYOL-A representations and Isolation Forest to energy-based audio segments, with no equations, fitted parameters, or self-citations that reduce any claimed result to the inputs by construction. Performance comparisons are made against external baselines on held-out multilingual datasets, keeping the derivation self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that laughter manifests as detectable energy anomalies in audio across languages and that BYOL-A embeddings preserve this property without supervision. No free parameters or invented entities are mentioned in the abstract.

axioms (1)
  • domain assumption Laughter can be reliably isolated as an energy-based anomaly in short audio segments without language-specific supervision.
    Invoked when the task is set up as anomaly detection on energy-segmented sequences.

pith-pipeline@v0.9.0 · 5455 in / 1173 out tokens · 24986 ms · 2026-05-14T21:14:56.673981+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

41 extracted references · 41 canonical work pages · 3 internal anchors

  1. [1]

    It is inherently social, as it not only communicates one’s internal state but also helps to propa- gate this state to other listeners [3]

    Introduction Laughter is ever-present in human interactions, playing an im- portant part in human-human communication, acting also as a tool for social bonding [1][2]. It is inherently social, as it not only communicates one’s internal state but also helps to propa- gate this state to other listeners [3]. It can express joy, relief, or success, but also a...

    work page

  2. [2]

    MultiLinguahah : A New Unsupervised Multilingual Acoustic Laughter Segmentation Method

    MultiLinguahah : Acoustic Laughter Segmentation The proposed method is composed of several steps. An overview is shown in Figure 1. 2.1. V oice Removal The first step of our approach consists of removing the speech from the audio signal, in order to retain the background, includ- ing laughter, music, and environmental sounds. In order to isolate the human...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  3. [3]

    Experiments and Results 3.1. Datasets for Evaluation We are validating and comparing models on a selection of 4 datasets containing laughter from various domains (in-the-wild, studio-recorded, and artificially created). StandUp4AI[28] dataset consists of 3,617 stand-up comedy videos spanning 7 languages. It includes audience laughter annotations, capturin...

    work page 2080

  4. [4]

    perform very similarly, with BYOL-A obtaining a slightly higher F1 at IoU=0.3, while wav2clip is marginally better at IoU=0.7. On TV Shows and YouTube, BYOL-A clearly out- performs wav2clip at both overlap thresholds, suggesting that self-supervised audio representations transfer particularly well to TV show data. Dataset Encoder F1 IoU=0.3 IoU=0.7 Stand-...

    work page

  5. [5]

    By combining a BYOL-A audio encoder with an Iso- lation Forest, our approach requires no labeled data and gener- alizes across languages and domains

    Conclusion We introduced MultiLinguahah, an unsupervised multilingual method for acoustic laughter segmentation that frames the task as anomaly detection over energy-based segmented audio se- quences. By combining a BYOL-A audio encoder with an Iso- lation Forest, our approach requires no labeled data and gener- alizes across languages and domains. Our ex...

    work page

  6. [6]

    Social laughter is correlated with an elevated pain threshold,

    R. I. M. Dunbar, R. Baron, A. Frangou, E. Pearce, E. J. C. Van Leeuwen, J. Stow, G. Partridge, I. MacDonald, V . Barra, and M. Van Vugt, “Social laughter is correlated with an elevated pain threshold,”Proceedings of the Royal Society B: Biological Sci- ences, vol. 279, no. 1731, pp. 1161–1167, 2012

    work page 2012

  7. [7]

    Laughter among deaf signers,

    R. R. Provine and K. Emmorey, “Laughter among deaf signers,” Journal of Deaf Studies and Deaf Education, vol. 11, no. 4, pp. 403–409, 2006

    work page 2006

  8. [8]

    The social psychology of humor,

    R. A. MARTIN, “The social psychology of humor,”The Psychol- ogy of Humor: An integrative approach, pp. 1–208, 2007

    work page 2007

  9. [9]

    Glenn,Laughter in interaction

    P. Glenn,Laughter in interaction. Cambridge University Press, 2003, vol. 18

    work page 2003

  10. [10]

    Semantic similarity of social functional smiles and laughter,

    A. Wood, S. Sievert, and J. Martin, “Semantic similarity of social functional smiles and laughter,”Journal of Nonverbal Behavior, vol. 46, no. 4, 2022

    work page 2022

  11. [11]

    Laughter as language,

    J. Ginzburg, C. Mazzocconi, and Y . Tian, “Laughter as language,” Glossa: a journal of general linguistics, vol. 5, no. 1, 2020

    work page 2020

  12. [12]

    Laughter research: a review of the ilhaire project,

    S. Dupont, H. C ¸ akmak, W. Curran, T. Dutoit, J. Hofmann, G. McKeown, O. Pietquin, T. Platt, W. Ruch, and J. Urbain, “Laughter research: a review of the ilhaire project,” inToward Robotic Socially Be-lievable Behaving Systems-Volume I: Model- ing Emo-tions, 2016, pp. 147–181

    work page 2016

  13. [13]

    SMILE: Multimodal Dataset for Understanding Laughter with Language Models,

    L. Hyun, K. Sung-Bin, S. Han, Y . Yu, and T. H. Oh, “SMILE: Multimodal Dataset for Understanding Laughter with Language Models,”Findings of the Association for Computational Linguis- tics: NAACL 2024 - Findings, pp. 1149–1167, 2024

    work page 2024

  14. [14]

    Lugrin, C

    B. Lugrin, C. Pelachaud, and D. Traum,The handbook on so- cially interactive agents: 20 years of research on embodied con- versational agents, intelligent virtual agents, and social robotics volume 2: interactivity, platforms, application. ACM, 2022

    work page 2022

  15. [15]

    UR- FUNNY: A Multimodal Language Dataset for Understanding Humor,

    M. K. Hasan, W. Rahman, A. Zadeh, J. Zhong, M. I. Tanveer, L.-P. Morency, Mohammed, and Hoque, “UR- FUNNY: A Multimodal Language Dataset for Understanding Humor,” inEMNLP-IJCNLP, 2019. [Online]. Available: http: //arxiv.org/abs/1904.06618

    work page arXiv 2019

  16. [16]

    Laugh- ter synthesis using pseudo phonetic tokens with a large-scale in- the-wild laughter corpus,

    D. Xin, S. Takamichi, A. Morimatsu, and H. Saruwatari, “Laugh- ter synthesis using pseudo phonetic tokens with a large-scale in- the-wild laughter corpus,” inProc. Interspeech, 2023

    work page 2023

  17. [17]

    Laughter and culture,

    G. A. Bryant and C. M. Bainbridge, “Laughter and culture,” Philosophical Transactions of the Royal Society B, vol. 377, no. 1863, p. 20210179, 2022

    work page 2022

  18. [18]

    Robust Laughter Segmenta- tion with Automatic Diverse Data Synthesis,

    T. Omine, K. Akita, and R. Tsuruno, “Robust Laughter Segmenta- tion with Automatic Diverse Data Synthesis,” inInterspeech, no. September, 2024, pp. 4748–4752

    work page 2024

  19. [19]

    Robust Laugh- ter Detection in Noisy Environments,

    J. Gillick, W. Deng, K. Ryokai, and D. Bamman, “Robust Laugh- ter Detection in Noisy Environments,” inProceedings of the An- nual Conference of the International Speech Communication As- sociation, INTERSPEECH, vol. 1. International Speech Com- munication Association, 2021, pp. 736–740

    work page 2021

  20. [20]

    Detection of laughter and screaming using the attention and ctc models,

    T. Matsuda and Y . Arimoto, “Detection of laughter and screaming using the attention and ctc models,” inProceedings of INTER- SPEECH, 2023, pp. 1025–1029

    work page 2023

  21. [21]

    Having Beer after Prayer? Measuring Cultural Bias in Large Language Models,

    T. Naous, M. J. Ryan, A. Ritter, and W. Xu, “Having Beer after Prayer? Measuring Cultural Bias in Large Language Models,” ACL, 2024. [Online]. Available: http://arxiv.org/abs/2305.14456

    work page arXiv 2024

  22. [22]

    Adapting Bias Evaluation to Domain Contexts using Generative Models,

    T. Quiroga, F. Bravo-Marquez, and V . Barriere, “Adapting Bias Evaluation to Domain Contexts using Generative Models,” inPro- ceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, C. Christodoulopoulos, T. Chakraborty, C. Rose, and V . Peng, Eds. Suzhou, China: Association for Computational Linguistics, 11 2025, pp. 28 055–2...

    work page 2025

  23. [23]

    A Study of Nationality Bias in Names and Perplexity using Off-the-Shelf Affect-related Tweet Classifiers,

    V . Barriere and S. Cifuentes, “A Study of Nationality Bias in Names and Perplexity using Off-the-Shelf Affect-related Tweet Classifiers,” inProceedings of EMNLP, 2024. [Online]. Available: https://aclanthology.org/2024.emnlp-main.34

    work page 2024

  24. [24]

    FunnyNet- W: Multimodal Learning of Funny Moments in Videos in the Wild,

    Z. S. Liu, R. Courant, and V . Kalogeiton, “FunnyNet- W: Multimodal Learning of Funny Moments in Videos in the Wild,”International Journal of Computer Vision, vol. 132, no. 8, pp. 2885–2906, 2024. [Online]. Available: https://doi.org/10.1007/s11263-024-02000-2

    work page doi:10.1007/s11263-024-02000-2 2024

  25. [25]

    Capturing, representing, and interacting with laughter,

    K. Ryokai, E. L ´opez, N. Howell, J. Gillick, and D. Bamman, “Capturing, representing, and interacting with laughter,” 04 2018, pp. 1–12

    work page 2018

  26. [26]

    Densely Connected Convolutional Networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger, “Densely Connected Convolutional Networks,” inCVPR, 2017

    work page 2017

  27. [27]

    Multi-Scale multi-band densenets for audio source separation,

    N. Takahashi and Y . Mitsufuji, “Multi-Scale multi-band densenets for audio source separation,” inIEEE Workshop on Applications of Signal Processing to Audio and Acoustics, vol. 2017-Octob, 2017, pp. 21–25

    work page 2017

  28. [28]

    BYOL for Audio: Exploring Pre-Trained General-Purpose Au- dio Representations,

    D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “BYOL for Audio: Exploring Pre-Trained General-Purpose Au- dio Representations,”IEEE/ACM Transactions on Audio Speech and Language Processing, vol. 31, pp. 137–151, 2023

    work page 2023

  29. [29]

    A Simple Method to Enhance Pre-trained Language Models with Speech Tokens for Classification

    N. Calbucura, J. Guillen, and V . Barriere, “A Simple Method to Enhance Pre-trained Language Models with Speech Tokens for Classification,” 4 2026. [Online]. Available: http: //arxiv.org/abs/2512.07571

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  30. [30]

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

    J. F. Gemmeke, D. P. W. 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), 2017, pp. 776–780

    work page 2017

  31. [31]

    Fsd50k: An open dataset of human-labeled sound events,

    E. Fonseca, X. Favory, J. Pons, F. Font, and X. Serra, “Fsd50k: An open dataset of human-labeled sound events,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 30, p. 829–852, Dec

    work page

  32. [32]

    Available: https://doi.org/10.1109/TASLP.2021

    [Online]. Available: https://doi.org/10.1109/TASLP.2021. 3133208

    work page doi:10.1109/taslp.2021 2021

  33. [33]

    Isolation forest,

    F. T. Liu, K. M. Ting, and Z.-H. Zhou, “Isolation forest,” in2008 Eighth IEEE International Conference on Data Mining, 2008, pp. 413–422

    work page 2008

  34. [34]

    Standup4ai: A new multilingual dataset for humor detection in stand-up comedy videos,

    V . Barriere, N. Gomez, L. Hemamou, S. Callejas, and B. Ravenet, “Standup4ai: A new multilingual dataset for humor detection in stand-up comedy videos,” inFindings of the Association for Com- putational Linguistics: EMNLP 2025. Suzhou, China: Asso- ciation for Computational Linguistics, Nov. 2025, pp. 16 951– 16 959

    work page 2025

  35. [35]

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

    J. F. Gemmeke, D. P. W. 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), 2017, pp. 776–780

    work page 2017

  36. [36]

    Face , Body , V oice : Video Person-Clustering with Multiple Modalities,

    A. Brown, V . Kalogeiton, and A. Zisserman, “Face , Body , V oice : Video Person-Clustering with Multiple Modalities,”ICCV Work- shops, pp. 3184–3194, 2021

    work page 2021

  37. [37]

    Multilingual Multimodal Detection of Humour in Stand-Up Comedy,

    A. Kuznetsova, “Multilingual Multimodal Detection of Humour in Stand-Up Comedy,” Ph.D. dissertation, 2024. [Online]. Available: https://aclanthology.org/2024.lrec-main.1037/

    work page 2024

  38. [38]

    100,000 Podcasts: A Spoken English Document Corpus,

    A. Clifton, S. Reddy, Y . Yu, A. Pappu, R. Rezapour, H. Bonab, M. Eskevich, G. J. Jones, J. Karlgren, B. Carterette, and R. Jones, “100,000 Podcasts: A Spoken English Document Corpus,”COL- ING 2020 - 28th International Conference on Computational Lin- guistics, Proceedings of the Conference, pp. 5903–5917, 2020

    work page 2020

  39. [39]

    V ocalsound: a Dataset for Improv- ing Human V ocal Sounds Recognition,

    Y . Gong, J. Yu, and J. Glass, “V ocalsound: a Dataset for Improv- ing Human V ocal Sounds Recognition,”ICASSP , IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing - Proceedings, vol. 2022-May, pp. 151–155, 2022

    work page 2022

  40. [40]

    Scikit-learn: Machine Learning in Python

    F. Pedregosa, G. Varoquaux, A. Gramfort, V . Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V . Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and ´E. Duchesnay, “Scikit-learn: Machine Learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825– 2830, 2012. [Online]. Available: http://dl.ac...

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  41. [41]

    Wav2Clip: Learning Robust Audio Representations From Clip,

    H. H. Wu, P. Seetharaman, K. Kumar, and J. P. Bello, “Wav2Clip: Learning Robust Audio Representations From Clip,”ICASSP , IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings, vol. 2022-May, pp. 4563–4567, 2022

    work page 2022