CoughViT: A Self-Supervised Vision Transformer for Cough Audio Representation Learning

Flora D. Salim; Hao Xue; Justin Luong

REVIEW 5 major objections 6 minor 50 references

CoughViT: A Self-Supervised Vision Transformer for Cough Audio Representation Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Self-supervised cough audio model beats supervised baselines

desk verdict A clean domain-specific self-supervised pretraining bake-off for cough audio, but the headline 'match or exceed SOTA' claim is not yet supported by the statistics. read the letter →

arxiv 2508.03764 v1 pith:Q7SFDPI7 submitted 2025-08-04 cs.SD cs.AIeess.AS

Justin Luong , Hao Xue , Flora D. Salim This is my paper · ORCID

classification cs.SDcs.AIeess.AS

keywords audioanalysiscoughclassificationself-supervisedlearningvisiontransformermaskedautoencoderspectrogramrespiratorydiseasediagnosistransfer

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that a vision transformer pre-trained by reconstructing masked patches of cough spectrograms learns general-purpose cough representations that transfer to diagnostic tasks with limited data. The authors train their model, CoughViT, on the unlabelled cough audio of the crowd-sourced COVID-19 Sounds dataset, using a masked autoencoding objective, and then fine-tune it on three tasks: COVID-19 detection, wet-or-dry cough classification, and cough detection. Compared against the same architecture with no pre-training, self-supervised pre-training lifts AUROC substantially; compared against a supervised AudioSet-pretrained audio transformer, CoughViT wins on COVID-19 detection and on a blind wet-or-dry test, while staying close on the other two tasks. The significance is that strong representations for cough analysis can be obtained without expensive clinical labels.

What carries the argument

The central mechanism is masked data modelling on log-mel spectrograms: a Vision Transformer (ViT-B) encoder is trained to reconstruct the pixel values of 16×16 patches of cough spectrograms after 75% of them are randomly masked, with the loss computed only on masked patches after patch normalisation. The masked patches are excluded from the encoder and reinserted as learnable mask tokens before a decoder, an efficiency device borrowed from masked autoencoders that leaves the encoder free to build a compact, reusable representation. Fixed sinusoidal positional encodings are used so the encoder can in principle accept spectrograms of different lengths, which matters because cough recordings are not normally of uniform duration. The pre-training is domain-specific: it uses only cough audio from COVID-19 Sounds, so the model learns cough-specific acoustic structure rather than general audio structure.

What would settle it

A controlled transfer experiment on a labelled dataset of non-COVID respiratory coughs (asthma, bronchitis, COPD, pneumonia) with clinically validated labels would settle the claim; if a CoughViT pre-trained on COVID-19 Sounds fine-tuned to those tasks performs no better than, or worse than, a model pre-trained on general audio or a well-engineered feature baseline, the representations are not general-purpose. A second, more targeted check is whether the fixed sinusoidal positional encodings actually handle spectrogram lengths unseen in pre-training without retraining, which the paper asserts but does not demonstrate.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that masked spectrogram reconstruction on in-domain cough audio produces representations that match or exceed the current state of the art in cough classification without using any labels during pre-training. Concretely, the pre-trained CoughViT encoder reaches 73.21 AUROC for COVID-19 detection (best in Table 1, ahead of 70.63 for AST-AudioSet), 98.25 for cough detection, and 74.95 for wet-or-dry classification (second, behind AST-AudioSet's 98.73 and 78.95). On the withheld COUGHVID blind test it scores 0.71 AUROC versus 0.56 for AST-AudioSet and 0.59 for a logistic regression baseline. In contrast, supervised pre-training on the same COVID-19 Sounds data with its self-reported labels gave little benefit or hurt downstream performance, supporting the paper's claim that self-supervision is the effective ingredient.

Load-bearing premise

The claim of general-purpose representations rests on the assumption that the coughs in COVID-19 Sounds, crowd-sourced largely during the pandemic with self-reported health labels, are acoustically representative enough of coughs in other respiratory conditions and recording environments to transfer to new tasks.

Editorial extensions

If this is right

Self-supervised pre-training on unlabelled cough audio is a viable route to improved diagnostic classifiers when annotations are scarce.
Pre-training on in-domain cough data can beat pre-training on large general audio datasets for some tasks, as shown by the COVID-19 detection and COUGHVID blind-test results.
Supervised pre-training using self-reported labels can degrade downstream transfer, so label quality matters as much as label quantity.
Non-overlapping spectrogram patches give performance comparable to overlapping-patch transformers with a smaller computational footprint.
A vision transformer can serve as the backbone for cough classification, with fixed sinusoidal positional encodings offering a simple route to variable input lengths.

Reading between the lines

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

Inference: the blind-test advantage on COUGHVID (0.71 vs 0.56) suggests that in-domain self-supervised features may generalise better to unseen recording conditions than general-audio features, but the paper only evaluates on a few datasets; a systematic out-of-distribution study would be needed to confirm this.
Inference: the same masked-autoencoder recipe could be applied to other scarce respiratory sound tasks such as wheeze detection, asthma severity, or paediatric pneumonia screening, since no labels are required for pre-training; the paper does not report such experiments.
Inference: it is unclear whether the learned representations encode disease-relevant acoustic cues or recording-device and environment invariants; comparing CoughViT against an augmentation-based contrastive pre-training objective would help separate these factors.
Inference: on the Edge-AI cough-detection blind test, CoughViT trails AST-AudioSet on both F1 scores, and the paper's sliding-window adaptation may not be tuned; a task-specific segmentation head rather than a binary classifier might close that gap.

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.

Desk Editor's Note

read the letter

Colleague,

Read CoughViT. The core claim—that a MAE-style ViT pretrained self-supervised on COVID-19 cough audio can match or beat an AudioSet-pretrained AST on downstream cough tasks—is plausible but the paper doesn't yet prove it. The experiments are clean in design: three tasks, a no-pretraining baseline, supervised domain-specific pretraining, self-supervised domain-specific pretraining, and AudioSet pretraining, plus an ablation over masking ratio, pooling, decoder attention, and pretraining epochs. That is a useful bake-off, and the blind test on COUGHVID and Edge-AI is more than most papers do.

What's actually new is narrow but real: the specific comparison of domain-specific MAE pretraining against broad supervised pretraining for cough audio, and the finding that the self-supervised approach competes with AudioSet pretraining on COVID-19 detection and cough detection. The writing is clear, and the claims are not egregiously overblown beyond the abstract.

Now the soft spots, in order of importance. First, no error bars or significance tests anywhere. The differences that drive the headline—2.58 AUROC on COVID-19 detection, 0.48 on cough detection, -4.0 on wet-or-dry—are within what could easily be noise across folds. Second, the reported configuration is the one that looked best after measuring AUROC on those exact benchmarks (Table 4 sweeps masking ratio, decoder attention, pooling, and epochs; the table's best row becomes the headline). That is selection on the test set, and the paper doesn't correct for it. Third, the COUGHVID blind test shows AST-AudioSet collapsing from 78.95 AUROC in Table 1 to 0.56 blind, while CoughViT goes from 74.95 to 0.71. That reversal is dramatic and unanalyzed; it suggests the Table 1 numbers are fragile to protocol changes. Fourth, the frame-count formula in Section 3.1.1 is wrong—should be (t-L)/H + 1, not (t-H)/L + 1. Minor, but it is a red flag. And the code isn't released, so the numbers can't be independently reproduced.

The central direction is sound. In-domain SSL pretraining clearly helps over no pretraining, and the paper's own baseline comparisons are honest. The 'first general-purpose cough representation' claim is a bit strong given prior self-supervised cough work, but they cite it. The 'match or exceed SOTA' phrasing needs to be softened to 'competitive, within noise' until the statistical grounding is fixed.

This deserves a serious referee. I'd send it out, but with a request for error bars, a more honest treatment of configuration selection, and a discussion of the blind test reversal. A solid conditional.

Recommendation: accept for peer review; expect heavy revision.

Referee Report

5 major / 6 minor

Summary. The paper proposes CoughViT, a self-supervised Vision Transformer pre-trained on cough audio spectrograms from the COVID-19 Sounds dataset via masked patch reconstruction. The resulting representations are fine-tuned on three downstream tasks: COVID-19 detection (Second DICOVA), wet-or-dry cough classification (COUGHVID), and cough detection (Edge-AI). The authors compare CoughViT against ViT and AST variants without pre-training, with supervised COVID-19 Sounds pre-training, and with AudioSet pre-training (AST-Audioset). They report that CoughViT achieves the best COVID-19 detection AUROC (73.21), is second to AST-Audioset on cough detection (98.25 vs 98.73) and wet-or-dry classification (74.95 vs 78.95), and outperforms AST-Audioset on the COUGHVID blind test (0.71 vs 0.56). The main claim is that domain-specific self-supervised pre-training matches or exceeds state-of-the-art supervised audio representations for cough tasks while avoiding label scarcity.

Significance. The paper addresses a relevant problem—label-efficient learning for cough audio analysis—and proposes a sensible approach: domain-specific self-supervised masked autoencoding on a large unlabeled cough corpus. The use of publicly available benchmarks and blind test sets, the systematic comparison with several baselines, and the ablation of architectural choices are strengths. If the reported results are statistically robust, the work could contribute to the development of adaptable, label-efficient respiratory sound classifiers. However, the central 'match or exceed' claim currently rests on small AUROC differences without uncertainty estimates, and the configuration of the main baseline is not entirely clear. With appropriate statistical validation and baseline clarification, the contribution would be valuable to the applied audio and health informatics community.

major comments (5)

[Section 4.4, Tables 1 and 4] The headline claim that CoughViT 'matches or exceeds' AST-Audioset is supported only by single AUROC values with no error bars, per-fold results, or significance tests. The reported differences (COVID-19 detection 73.21 vs 70.63, cough detection 98.25 vs 98.73, wet-or-dry 74.95 vs 78.95) are small, and the best CoughViT configuration was selected after evaluating 12 variants on the same three benchmarks (Table 4). This makes the observed advantage over AST-Audioset on COVID-19 detection indistinguishable from selection noise. Please report variance across folds, multiple seeds, or a nested validation scheme that separates configuration selection from evaluation.
[Section 4.3 and Table 1] The baseline 'AST Audioset' is ambiguous: the column 'Pre-training Epochs' reads 5, but it is unclear whether the authors fine-tuned an existing AST checkpoint pre-trained on AudioSet (as commonly released) or pre-trained AST on AudioSet themselves for only 5 epochs. Since AST-Audioset is the state-of-the-art comparator against which the central claim is made, this ambiguity is load-bearing. If the latter, the comparison may understate AST's performance; if the former, the column label is misleading. Please specify the exact source and pre-training setup of AST-Audioset.
[Section 3.1.1] The frame-count formula w = (t-H)/L + 1 is incorrect for frame length L and hop H; the standard formula is floor((t-L)/H) + 1. For example, with t=1s, L=25ms, H=10ms, the formula gives 40.6, which is not an integer. Additionally, the text claims ViTs handle varying input lengths by adjusting positional encodings, but the method immediately truncates or pads spectrograms to a fixed number of frames for batching, and no experiment demonstrates adaptation to different input lengths. This internal contradiction weakens the architectural motivation.
[Section 4.5 and Table 2] The COUGHVID blind test results reverse the ranking from Table 1 for wet-or-dry classification: CoughViT achieves 0.71 AUROC versus 0.56 for AST-Audioset, whereas Table 1 shows CoughViT at 74.95 and AST-Audioset at 78.95. The paper reports this reversal without analysis, leaving the reader to wonder whether the Table 1 difference is a result of overfitting to the public labels, differences in evaluation protocol, or another factor. Please investigate and explain this discrepancy, as it directly affects the 'match or exceed' claim.
[Abstract, Section 5] The term 'general-purpose cough representations' is stronger than the evidence presented. All three downstream tasks are either COVID-19 detection or cough characteristics (wet/dry, cough detection) evaluated on datasets collected largely during the pandemic; no evaluation on asthma, bronchitis, COPD, or other respiratory conditions is provided. The conclusion's acknowledgment of limited dataset variety is appropriate, but the abstract and introduction should be qualified to avoid overclaiming generality.

minor comments (6)

[Section 1] The list of contributions enumerates items (1) and (3) but omits item (2); please renumber or add the missing contribution.
[Section 4.3] In the sentence 'There serve as ways to evaluate...', 'There' should be 'These'.
[Section 4.5] The phrase 'we perform a evaluation' should be 'we perform an evaluation'.
[Section 3.2.1] The sentence 'Each dataset samples consists of...' should read 'Each dataset sample consists of...'.
[Equation (1)] The patch-count formula lacks floor or ceiling operators, which is important for non-divisible spectrogram dimensions; please clarify the rounding behavior.
[Table 4] The use of 'N/A' in the no-pre-training rows is not explained; consider using a dash or a footnote to indicate inapplicability.

Circularity Check

1 steps flagged · score 4.0 of 10

Core self-supervised representation learning is label-free and independent, but the headline 'match or exceed' claim is partly a selection artifact because the reported CoughViT configuration was chosen after measuring AUROC on the same three benchmark tasks.

fitted input called prediction [Section 4.3 (Models) and Section 4.6 (CoughViT Ablation), Table 4 caption]
"We explore a variety of configurations and compare the best performing ones to existing methods. ... Ablation Study of CoughViT Configurations. Compares AUROC(%) after fine-tuning on downstream tasks, with all pre-training conducted on the COVID-19 Sounds dataset. Variations included pre-training strategy and duration, masking ratio, self-attention type for spectrogram reconstruction, and classification method."

The paper's central claim that CoughViT representations 'match or exceed current state-of-the-art supervised audio representations' rests on the AUROC values in Table 1. But the CoughViT row is not a pre-specified model: it is the best of the 12 configurations evaluated on the same three downstream benchmarks in Table 4 (masking ratio 0.50 vs 0.75, global vs windowed decoder attention, class token vs mean pooling, 100 vs 200 epochs). Choosing the best row based on AUROC on exactly the tasks that are then used as evidence of superiority makes the comparison partially in-sample: the reported advantage over AST-Audioset is inflated by selection among many configurations rather than being an independent prediction of the method.

full rationale

The paper's central contribution is a self-supervised masked-spectrogram-reconstruction pretraining task on COVID-19 Sounds, which does not use downstream labels and is therefore not circular in its representation-learning objective. The pre-training loss is patch-normalized MSE, and the encoder weights are learned from unlabeled cough audio. The downstream evaluation uses three separate public benchmarks, and the authors also report two blind test sets (COUGHVID and Edge-AI), which provide some out-of-sample validation independent of the authors' own tuning. However, the quantitative claim that CoughViT 'matches or exceeds' supervised state-of-the-art is weakened by the configuration-selection procedure: Section 4.3 says the authors compare 'the best performing ones' from a set of configurations, and Table 4 scores all configurations on the same downstream tasks used for the headline comparison. The final model is therefore chosen after inspecting the evaluation metrics, so the reported AUROC gaps (e.g., 73.21 vs 70.63 on COVID-19 detection) are partly the product of model selection on those very benchmarks. This is a real but moderate circularity concern; it does not reduce the representation learning itself to its inputs. There is also a self-citation ([46], Xue and Salim) but it is descriptive related work and is not load-bearing for the paper's main argument, so it does not raise the score further.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

No new physical or conceptual entities are introduced. The central claim rests on standard machine-learning assumptions about spectrogram representations and transfer learning, plus the domain-specific assumption that COVID-19 Sounds cough audio generalizes to other cough types. Key hyperparameters (masking ratio, patch size, epochs, decoder width) are inherited from prior work or chosen by hand/ablation.

free parameters (5)

masking_ratio = 0.75
Selected after comparing 0.50 and 0.75 on downstream AUROC for the same three benchmarks (Table 4); 0.75 chosen as best.
pretraining_epochs = 100
Chosen for pragmatic reasons; validation loss still dropping at cutoff (Section 4.6). Additional 100 epochs gave mixed results.
patch_size = 16x16
Standard ViT/MAE patch size, adopted without domain-specific tuning (Section 3.1.2).
decoder_hidden_dim = 524 (likely typo for 512)
Section 3.2.3 states 524-dimension vectors; this is inconsistent with standard MAE decoder widths and appears to be a typo.
spectrogram_parameters = 128 mel bins, 25ms frame, 10ms hop
Chosen based on common audio preprocessing conventions; not optimized for cough.

assumptions (6)

domain assumption Log-mel spectrograms are an adequate representation of cough audio for classification.
Used throughout; all models operate on log-mel spectrograms (Section 3.1.1).
domain assumption Masked patch reconstruction on spectrograms learns transferable representations for downstream classification.
Borrowed from MAE/Audio-MAE; central to the pretraining method (Section 3.2.3).
domain assumption The COVID-19 Sounds dataset is representative enough of cough audio to serve as general-purpose pretraining data.
This is load-bearing for the 'general-purpose' claim (Sections 3.2.1 and 5).
domain assumption Self-reported labels in the benchmark datasets are reliable enough for evaluating diagnostic performance.
COVID-19 detection and wet-or-dry labels rely on self-report or expert annotations with known reliability issues (Sections 4.1).
domain assumption AUROC with 5-fold cross-validation is an appropriate measure of model performance for these tasks.
Used as the primary metric for all tasks (Section 4.2).
ad hoc to paper Fixed sinusoidal positional encodings can be adapted to arbitrary input sequence lengths without retraining.
The paper asserts that ViTs handle varying input lengths by adjusting positional encodings, but does not specify or validate the interpolation method (Section 3.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoughViT: A Self-Supervised Vision Transformer for Cough Audio Representation Learning." pith.science (2026). https://pith.science/paper/Q7SFDPI7

@misc{pith2026250803764,
  author       = {Pith},
  title        = {Pith review of: CoughViT: A Self-Supervised Vision Transformer for Cough Audio Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q7SFDPI7}},
  note         = {Machine review of arXiv:2508.03764}
}

read the original abstract

Physicians routinely assess respiratory sounds during the diagnostic process, providing insight into the condition of a patient's airways. In recent years, AI-based diagnostic systems operating on respiratory sounds, have demonstrated success in respiratory disease detection. These systems represent a crucial advancement in early and accessible diagnosis which is essential for timely treatment. However, label and data scarcity remain key challenges, especially for conditions beyond COVID-19, limiting diagnostic performance and reliable evaluation. In this paper, we propose CoughViT, a novel pre-training framework for learning general-purpose cough sound representations, to enhance diagnostic performance in tasks with limited data. To address label scarcity, we employ masked data modelling to train a feature encoder in a self-supervised learning manner. We evaluate our approach against other pre-training strategies on three diagnostically important cough classification tasks. Experimental results show that our representations match or exceed current state-of-the-art supervised audio representations in enhancing performance on downstream tasks.

Figures

Figures reproduced from arXiv: 2508.03764 by the authors.

**Figure 1.** Overview of the Vision Transformer architecture for masked data modelling. A random portion of patches are masked [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

50 extracted references · 38 canonical work pages

[1]

Kawther S Alqudaihi, Nida Aslam, Irfan Ullah Khan, Abdullah M Almuhaideb, Shikah J Alsunaidi, Nehad M Abdel Rahman Ibrahim, Fahd A Alhaidari, Fatema S Shaikh, Yasmine M Alsenbel, Dima M Alalharith, et al . 2021. Cough sound detection and diagnosis using artificial intelligence techniques: challenges and opportunities. Ieee Access 9 (2021), 102327–102344

2021
[2]

Yusuf Amrulloh, Udantha Abeyratne, Vinayak Swarnkar, and Rina Triasih. 2015. Cough sound analysis for pneumonia and asthma classification in pediatric population. In 2015 6th international conference on intelligent systems, modelling and simulation. IEEE, 127–131

2015
[3]

Yusuf A Amrulloh, Udantha R Abeyratne, Vinayak Swarnkar, Rina Triasih, and Amalia Setyati. 2015. Automatic cough segmentation from non-contact sound recordings in pediatric wards. Biomedical Signal Processing and Control 21 (2015), 126–136

2015
[4]

Filipe Barata, David Cleres, Peter Tinschert, Chen-Hsuan Iris Shih, Frank Ras- souli, Maximilian Boesch, Martin Brutsche, Elgar Fleisch, et al. 2023. Nighttime Continuous Contactless Smartphone-Based Cough Monitoring for the Ward: Validation Study. JMIR Formative Research 7, 1 (2023), e38439. CoughViT: A Self-Supervised Vision Transformer for Cough Audio ...

work page 2023
[5]

SS Birring, T Fleming, S Matos, AA Raj, DH Evans, and ID Pavord. 2008. The Le- icester Cough Monitor: preliminary validation of an automated cough detection system in chronic cough. European Respiratory Journal 31, 5 (2008), 1013–1018

work page 2008
[6]

Chloë Brown, Jagmohan Chauhan, Andreas Grammenos, Jing Han, Apinan Hasthanasombat, Dimitris Spathis, Tong Xia, Pietro Cicuta, and Cecilia Mascolo

work page
[7]

Jobie Budd, Kieran Baker, Emma Karoune, Harry Coppock, Selina Patel, Ana Ten- dero Cañadas, Alexander Titcomb, Richard Payne, David Hurley, Sabrina Eggle- stone, et al . 2022. A large-scale and PCR-referenced vocal audio dataset for COVID-19. arXiv preprint arXiv:2212.07738 (2022)

work page Pith review arXiv 2022
[8]

Edresson Casanova, Arnaldo Candido, Ricardo Corso Fernandes, Marcelo Fin- ger, Lucas Rafael Stefanel Gris, Moacir Antonelli Ponti, Daniel Peixoto Pinto Da Silva, et al. 2021. Transfer learning and data augmentation techniques to the COVID-19 identification tasks in ComParE 2021. In22nd Annual Conference of the International Speech Communication Associatio...

work page 2021

Show all 50 references

[9]

Scott Claxton, Paul Porter, Joanna Brisbane, Natasha Bear, Javan Wood, Vesa Peltonen, Phillip Della, Claire Smith, and Udantha Abeyratne. 2021. Identifying acute exacerbations of chronic obstructive pulmonary disease using patient- reported symptoms and cough feature analysis....

2021
[10]

Harry Coppock, Lyn Jones, Ivan Kiskin, and Björn Schuller. 2021. COVID-19 detection from audio: seven grains of salt. The Lancet Digital Health 3, 9 (2021), e537–e538

2021
[11]

Ting Dang, Thomas Quinnell, and Cecilia Mascolo. 2022. Exploring Semi- supervised Learning for Audio-based COVID-19 Detection using FixMatch. Proc. Interspeech 2022 (2022), 2468–2472

2022
[12]

Vladimir Despotovic, Muhannad Ismael, Maël Cornil, Roderick Mc Call, and Guy Fagherazzi. 2021. Detection of COVID-19 from voice, cough and breathing patterns: Dataset and preliminary results. Computers in Biology and Medicine 138 (2021), 104944

2021
[13]

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

2018 arXiv
[14]

Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv prepri...

2020 arXiv
[15]

Yutong Feng, Jianwen Jiang, Mingqian Tang, Rong Jin, and Yue Gao. 2021. Re- thinking supervised pre-training for better downstream transferring. arXiv preprint arXiv:2110.06014 (2021)

2021 arXiv
[16]

Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. Audio set: An ontology and human-labeled dataset for audio events. In2017 IEEE international conference on acoustics, speech and signal process...

2017
[17]

Yuan Gong, Yu-An Chung, and James Glass. 2021. Ast: Audio spectrogram transformer. arXiv preprint arXiv:2104.01778 (2021)

2021 arXiv
[18]

Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass. 2022. Ssast: Self- supervised audio spectrogram transformer. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 36. 10699–10709

2022
[19]

Yu Gu, Robert Tinn, Hao Cheng, Michael Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2021. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Computing for Healthcare (HEALTH) 3, 1 (...

2021
[20]

Arati Gurung, Carolyn G Scrafford, James M Tielsch, Orin S Levine, and William Checkley. 2011. Computerized lung sound analysis as diagnostic aid for the detection of abnormal lung sounds: a systematic review and meta-analysis. Respiratory medicine 105, 9 (2011), 1396–1403

2011
[21]

Honorata Hafke-Dys, Anna Bręborowicz, Paweł Kleka, Jędrzej Kociński, and Adam Biniakowski. 2019. The accuracy of lung auscultation in the practice of physicians and medical students. PLoS One 14, 8 (2019), e0220606

2019
[22]

Jing Han, Tong Xia, Dimitris Spathis, Erika Bondareva, Chloë Brown, Jagmohan Chauhan, Ting Dang, Andreas Grammenos, Apinan Hasthanasombat, Andres Floto, et al . 2022. Sounds of COVID-19: exploring realistic performance of audio-based digital testing. NPJ digital medicine 5, 1 ...

2022
[23]

Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick
[24]

Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. 2017. CNN architectures for large-scale audio classification. In2017 ieee international conference on acoustics, sp...

2017
[25]

Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer. 2022. Masked autoencoders that listen. Advances in Neural Information Processing Systems 35 (2022), 28708– 28720

2022
[26]

Aneeqa Ijaz, Muhammad Nabeel, Usama Masood, Tahir Mahmood, Mydah Sajid Hashmi, Iryna Posokhova, Ali Rizwan, and Ali Imran. 2022. Towards using cough for respiratory disease diagnosis by leveraging Artificial Intelligence: A survey. Informatics in Medicine Unlocked 29 (2022), 100832

2022
[27]

Ali Imran, Iryna Posokhova, Haneya N Qureshi, Usama Masood, Muham- mad Sajid Riaz, Kamran Ali, Charles N John, MD Iftikhar Hussain, and Muham- mad Nabeel. 2020. AI4COVID-19: AI enabled preliminary diagnosis for COVID- 19 from cough samples via an app. Informatics in medicine u...

2020
[28]

Yoonjoo Kim, YunKyong Hyon, Sunju Lee, Seong-Dae Woo, Taeyoung Ha, and Chaeuk Chung. 2022. The coming era of a new auscultation system for analyzing respiratory sounds. BMC Pulmonary Medicine 22, 1 (2022), 119

2022
[29]

Andy T Liu, Shu-wen Yang, Po-Han Chi, Po-chun Hsu, and Hung-yi Lee. 2020. Mockingjay: Unsupervised speech representation learning with deep bidirec- tional transformer encoders. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASS...

2020
[30]

Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision. 10012–10022

2021
[31]

Igor DS Miranda, Andreas H Diacon, and Thomas R Niesler. 2019. A comparative study of features for acoustic cough detection using deep architectures. In 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 2601–2605

2019
[32]

Lara Orlandic, Tomas Teijeiro, and David Atienza. 2021. The COUGHVID crowd- sourcing dataset, a corpus for the study of large-scale cough analysis algorithms. Scientific Data 8, 1 (2021), 156

2021
[33]

Lara Orlandic, Tomas Teijeiro, and David Atienza. 2023. A semi-supervised algorithm for improving the consistency of crowdsourced datasets: The COVID- 19 case study on respiratory disorder classification. Computer Methods and Programs in Biomedicine 241 (2023), 107743

2023
[34]

Lara Orlandic, Jérôme Paul Rémy Thevenot, Tomas Teijeiro, and David Atienza Alonso. 2023. A Multimodal Dataset for Automatic Edge-AI Cough Detection. In The 45th Annual International Conference of the IEEE Engineering in Medicine and Biology Society

2023
[35]

Madhurananda Pahar, Marisa Klopper, Robin Warren, and Thomas Niesler. 2021. COVID-19 cough classification using machine learning and global smartphone recordings. Computers in Biology and Medicine 135 (2021), 104572

2021
[36]

Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. 2015. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 5206–5210

2015
[37]

Gadi Pinkas, Yarden Karny, Aviad Malachi, Galia Barkai, Gideon Bachar, and Vered Aharonson. 2020. SARS-CoV-2 detection from voice. IEEE Open Journal of Engineering in Medicine and Biology 1 (2020), 268–274

2020
[38]

Paul Porter, Udantha Abeyratne, Vinayak Swarnkar, Jamie Tan, Ti-wan Ng, Joanna M Brisbane, Deirdre Speldewinde, Jennifer Choveaux, Roneel Sharan, Keegan Kosasih, et al. 2019. A prospective multicentre study testing the diag- nostic accuracy of an automated cough sound centred ...

2019
[39]

Hendrik Purwins, Bo Li, Tuomas Virtanen, Jan Schlüter, Shuo-Yiin Chang, and Tara Sainath. 2019. Deep learning for audio signal processing. IEEE Journal of Selected Topics in Signal Processing 13, 2 (2019), 206–219

2019
[40]

Malay Sarkar, Irappa Madabhavi, Narasimhalu Niranjan, and Megha Dogra. 2015. Auscultation of the respiratory system. Annals of thoracic medicine 10, 3 (2015), 158

2015
[41]

Neeraj Sharma, Prashant Krishnan, Rohit Kumar, Shreyas Ramoji, Srikanth Raj Chetupalli, Prasanta Kumar Ghosh, Sriram Ganapathy, et al. 2020. Coswara–a database of breathing, cough, and voice sounds for COVID-19 diagnosis. arXiv preprint arXiv:2005.10548 (2020)

2020 arXiv
[42]

Neeraj Kumar Sharma, Srikanth Raj Chetupalli, Debarpan Bhattacharya, De- bottam Dutta, Pravin Mote, and Sriram Ganapathy. 2022. The second DICOVA challenge: Dataset and performance analysis for diagnosis of COVID-19 using acoustics. In ICASSP 2022-2022 IEEE International Confe...

2022
[43]

Vinayak Swarnkar, Udantha Abeyratne, Jamie Tan, Ti Wan Ng, Joanna M Bris- bane, Jennifer Choveaux, and Paul Porter. 2021. Stratifying asthma severity in children using cough sound analytic technology. Journal of Asthma 58, 2 (2021), 160–169

2021
[44]

Vinayak Swarnkar, Udantha R Abeyratne, Anne B Chang, Yusuf A Amrulloh, Amalia Setyati, and Rina Triasih. 2013. Automatic identification of wet and dry cough in pediatric patients with respiratory diseases. Annals of biomedical engineering 41 (2013), 1016–1028

2013
[45]

Tong Xia, Dimitris Spathis, J Ch, Andreas Grammenos, Jing Han, Apinan Hasthanasombat, Erika Bondareva, Ting Dang, Andres Floto, Pietro Cicuta, et al
[46]

Hao Xue and Flora D Salim. 2021. Exploring self-supervised representation ensembles for COVID-19 cough classification. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 1944–1952

2021
[47]

Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. 2022. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2736–2746

2022
[2020]

arXiv preprint arXiv:2006.05919 (2020)

Exploring automatic diagnosis of COVID-19 from crowdsourced respiratory sound data. arXiv preprint arXiv:2006.05919 (2020)

2020 arXiv
[2021]

In Thirty-fifth Conference on Neural Information Processing Systems Justin Luong, Hao Xue, and Flora D

COVID-19 Sounds: A Large-Scale Audio Dataset for Digital Respiratory Screening. In Thirty-fifth Conference on Neural Information Processing Systems Justin Luong, Hao Xue, and Flora D. Salim Datasets and Benchmarks Track (Round 2)
[2022]

In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16000–16009

Pith tools

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

[1] [1]

Kawther S Alqudaihi, Nida Aslam, Irfan Ullah Khan, Abdullah M Almuhaideb, Shikah J Alsunaidi, Nehad M Abdel Rahman Ibrahim, Fahd A Alhaidari, Fatema S Shaikh, Yasmine M Alsenbel, Dima M Alalharith, et al . 2021. Cough sound detection and diagnosis using artificial intelligence techniques: challenges and opportunities. Ieee Access 9 (2021), 102327–102344

2021

[2] [2]

Yusuf Amrulloh, Udantha Abeyratne, Vinayak Swarnkar, and Rina Triasih. 2015. Cough sound analysis for pneumonia and asthma classification in pediatric population. In 2015 6th international conference on intelligent systems, modelling and simulation. IEEE, 127–131

2015

[3] [3]

Yusuf A Amrulloh, Udantha R Abeyratne, Vinayak Swarnkar, Rina Triasih, and Amalia Setyati. 2015. Automatic cough segmentation from non-contact sound recordings in pediatric wards. Biomedical Signal Processing and Control 21 (2015), 126–136

2015

[4] [4]

Filipe Barata, David Cleres, Peter Tinschert, Chen-Hsuan Iris Shih, Frank Ras- souli, Maximilian Boesch, Martin Brutsche, Elgar Fleisch, et al. 2023. Nighttime Continuous Contactless Smartphone-Based Cough Monitoring for the Ward: Validation Study. JMIR Formative Research 7, 1 (2023), e38439. CoughViT: A Self-Supervised Vision Transformer for Cough Audio ...

work page 2023

[5] [5]

SS Birring, T Fleming, S Matos, AA Raj, DH Evans, and ID Pavord. 2008. The Le- icester Cough Monitor: preliminary validation of an automated cough detection system in chronic cough. European Respiratory Journal 31, 5 (2008), 1013–1018

work page 2008

[6] [6]

Chloë Brown, Jagmohan Chauhan, Andreas Grammenos, Jing Han, Apinan Hasthanasombat, Dimitris Spathis, Tong Xia, Pietro Cicuta, and Cecilia Mascolo

work page

[7] [7]

Jobie Budd, Kieran Baker, Emma Karoune, Harry Coppock, Selina Patel, Ana Ten- dero Cañadas, Alexander Titcomb, Richard Payne, David Hurley, Sabrina Eggle- stone, et al . 2022. A large-scale and PCR-referenced vocal audio dataset for COVID-19. arXiv preprint arXiv:2212.07738 (2022)

work page Pith review arXiv 2022

[8] [8]

Edresson Casanova, Arnaldo Candido, Ricardo Corso Fernandes, Marcelo Fin- ger, Lucas Rafael Stefanel Gris, Moacir Antonelli Ponti, Daniel Peixoto Pinto Da Silva, et al. 2021. Transfer learning and data augmentation techniques to the COVID-19 identification tasks in ComParE 2021. In22nd Annual Conference of the International Speech Communication Associatio...

work page 2021