Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Sound Signal Synthesis with Auxiliary Classifier GAN, COVID-19 cough as an example

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Adding ACGAN-generated synthetic cough spectrograms to the training set raises a COVID-19 cough classifier from 72% to 75% accuracy on held-out data.

desk verdict The 3-point gain is almost certainly an artifact of comparing a validation-set maximum to a test-set score; the paper is an honest but under-controlled feasibility study. read the letter →

arxiv 2508.08892 v1 pith:2BRE4QIJ submitted 2025-08-12 cs.SD cs.LG

Yahya Sherif Solayman Mohamed Saleh , Ahmed Mohammed Dabbous , Lama Alkhaled , Hum Yan Chai , Muhammad Ehsan Rana , Hamam Mokayed This is my paper · ORCID
classification cs.SDcs.LG
keywords ACGANCOVID-19coughaudioMelspectrogramdataaugmentationsignalprocessingclassificationscarcity
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 asks whether synthetic cough audio can relieve the data scarcity that limits COVID-19 detection from sound. The authors train a CNN on balanced, self-supervised-labeled Mel spectrograms from the COUGHVID dataset to a 72% baseline, train an Auxiliary Classifier GAN (ACGAN) to generate healthy and COVID-19 cough spectrograms, and retrain the classifier after adding 1,000 generated samples per class. They report that held-out classification accuracy rises to 75%. The study is a proof of concept: the authors flag noisy self-reported labels, visible artifacts in generated spectrograms, and only moderate audible realism, so the value is in showing that class-conditional synthesis can nudge a scarce-data audio classifier.

What carries the argument

The central object is the Auxiliary Classifier GAN (ACGAN), a GAN whose generator takes both latent noise and a class label and whose discriminator outputs both a real/fake verdict and a class-label distribution. That dual output is what forces the generator to learn features that separate healthy from COVID-19 cough spectrograms, and the label conditioning is what allows novel samples of the scarce COVID-19 class to be produced. A supporting piece is the signal-processing front end: cough audio is segmented, normalized, and converted to 128×24 Mel spectrograms, which gives the GAN and the CNN a common fixed-size image representation.

What would settle it

Re-run the pipeline and report the baseline CNN's accuracy on the exact test set used for the 75% result; if it is already 75% or higher, the gain disappears. As a second check, replace the synthetic samples with an equal number of real samples or standard augmentations; if accuracy improves just as much, the ACGAN is not the cause.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that label-conditioned synthetic spectrograms carry enough class-relevant structure to improve a cough-based healthy-versus-COVID classifier: accuracy moves from 72% to 75% on the held-out test partition of COUGHVID's status_SSL-labeled data. The paper does not claim the synthetic audio is medically faithful—the audible quality is moderate and the authors cannot confirm COVID-specific traces in the generated samples. The contribution is that ACGAN augmentation functions as a workable data-scarcity remedy: the generator learns to produce both classes conditionally, and the retrained CNN generalizes slightly better than the baseline on data it has not

Load-bearing premise

The claimed 3-point gain assumes the 72% baseline and the 75% augmented result are measured on the same held-out test set with the same checkpoint rule, and that the self-reported cough labels carry the signal the models learn.

Editorial extensions

If this is right

  • ACGAN-style augmentation becomes a candidate remedy whenever a diagnostic audio class is too rare to train on directly.
  • Conditional generation lets the training set be rebalanced toward the minority class without literally duplicating real samples.
  • The approach imports the mature tooling of image GANs into audio diagnosis by working on spectrograms instead of waveforms.
  • Because the reported gain is small, synthetic data is a complement to—not a replacement for—better labels and careful segmentation.

Reading between the lines

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

  • A direct way to test the mechanism: replace the synthetic spectrograms with the same number of real healthy/COVID spectrograms or standard audio augmentations; if accuracy rises equally, the ACGAN's synthetic content is not what is driving the gain.
  • Since the status_SSL labels are outputs of a self-supervised model rather than PCR-confirmed diagnoses, the 75% figure may measure consistency with that labeler, not detection of disease; expert-annotated or RT-PCR-confirmed evaluation would separate the two.
  • The same pipeline could be probed on audio conditions with stronger acoustic signatures—such as tuberculosis or croup—where the generated class signal may be easier to learn than in crowdsourced COVID coughs.
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

4 major / 6 minor

Summary. The paper proposes using an Auxiliary Classifier GAN (ACGAN) to synthesize COVID-19 cough Mel spectrograms and using these synthetic samples to augment the training set of a cough classifier. The authors train a CNN baseline on a balanced, cleaned subset of the COUGHVID dataset with status_SSL labels, report a baseline accuracy of 72%, then retrain an identical CNN on the training set augmented with ACGAN-generated Mel spectrograms and report a test accuracy of 75%. The paper also documents dataset biases, preprocessing, ACGAN architecture and training instability, and candidly discusses label noise and segmentation issues. The core claim is that class-conditional GAN augmentation yields a small but real improvement in held-out accuracy.

Significance. If substantiated, this would be a useful contribution: audio-domain COVID-19 cough synthesis has been largely unexplored relative to CXR/CT synthesis, and the COUGHVID-based experimental setup is a realistic testbed. The manuscript is also honest about confounding factors, especially label noise, and it makes the ACGAN design choices (instance noise, soft labels, label embedding) fairly explicit. However, the paper's empirical support is currently under-powered: the key 72%-to-75% comparison is not made under a consistent evaluation protocol, no uncertainty quantification is provided, and the attribution of the gain specifically to ACGAN augmentation is not tested against simpler augmentation controls. The claim is plausible but not yet established as reported.

major comments (4)
  1. [Section 7.1 and Section 7.2] The central comparison is not commensurate. Section 7.1 states the baseline 'reached a maximum value of 72% on the validation set' (Fig. 12), while Section 7.2 reports the augmented model 'evaluated on the testing set once more' and reaching 75%. The baseline's accuracy on the same held-out test set is never given, and the augmented model's 75% is not tied to the same checkpoint-selection rule (best-validation vs. final-epoch). Without this, the reported 3-point gain may be entirely due to protocol mismatch. Please report baseline test accuracy under the identical evaluation protocol and checkpoint rule.
  2. [Section 6 vs. Section 7.2] The augmentation count is inconsistent. Section 6 says '200 synthetic images ... from each class', corresponding to 25% of the initial training set, while Section 7.2 says '1000 novel samples from the COVID-19 and Healthy class labels' are generated and mixed into the training set. These are not the same number, and the discrepancy affects the comparison (dataset size and class balance change between baseline and augmented training). Please reconcile and specify per-class counts.
  3. [Section 7.2] No repeated runs, error bars, confidence intervals, or significance testing are reported. A 3-point accuracy difference on a single 80/10/10 split with a small, noisy dataset is within plausible seed-to-seed variance for CNNs. At minimum, report mean and std across several seeds for both the baseline and augmented models, and a significance test or effect-size statement.
  4. [Section 7.2] The causal attribution to ACGAN augmentation is not tested against simpler controls. Retraining on real samples duplicated to the same size, or using standard augmentation (noise, time-shift), would show whether the improvement comes from adding synthetic spectrograms specifically or merely from a larger training set/regularization. Currently the design cannot rule out the latter.
minor comments (6)
  1. [Figure 8] Figure 8(b) is labeled 'SIFT' but the text and context indicate STFT (Short-Time Fourier Transform).
  2. [Section 8] Discussion refers to 'synthesizing CET scans'; this should be 'CT scans'.
  3. [Section 3, Eq. (1)] The notation uses '∨' where conditional probability '|' is intended, and the subscripts on the loss expressions mix the optimization variables (e.g., L_D and L_G both show max over C). This makes the objective hard to parse and should be rewritten.
  4. [Section 7.1] The classifier architecture is described as 'matches the discriminator architecture without the validation layer', but elsewhere the discriminator output is called the 'validity layer'. Please use consistent terminology.
  5. [Section 7.2 / Figures 14-16] The qualitative assessment of synthetic samples (artifacts, energy distribution, 'columnar shape') is useful but subjective. Quantitative measures—e.g., FID, Inception Score, or spectrogram-level similarity—would strengthen the claim that the GAN output is usable.
  6. [Throughout] Reference formatting is inconsistent (e.g., missing years, duplicate entries for Haritaoglu et al., 'Ai et el.' typo, and the reference list contains entries not cited in the text). A thorough copyedit is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the ACGAN augmentation claim is evaluated on a held-out test set, and no prediction reduces by construction to its fitted inputs.

full rationale

The paper's derivation chain is: (1) train a baseline CNN on balanced Coughvid status_SSL-labeled Mel spectrograms; (2) train an ACGAN on the same training split; (3) generate class-conditional synthetic spectrograms; (4) retrain a CNN on the augmented training set; and (5) evaluate on an independent held-out testing set. Each stage uses only training-distribution information, and the final 75% figure is a measured test-set result rather than an algebraic consequence of the training objective. The synthetic samples are conditioned on class labels, not on test labels, so the improvement claim is not definitionally forced. The paper does cite its own prior work (Saleh et al. 2024, Mokayed et al. 2024), but these citations appear only as general literature context and are not load-bearing for the central claim. There is no imported uniqueness theorem, no ansatz smuggled via self-citation, and no renaming of a known result as a new contribution. The genuine weaknesses are validity concerns, not circularity: Section 7.1 reports the 72% baseline as a validation-set maximum, while Section 7.2 reports the 75% augmented result on the testing set, and the paper never reports the baseline CNN's accuracy on that same testing set. This makes the 3-point improvement difficult to verify but does not make it circular. Similarly, Section 9 candidly acknowledges noisy self-reported/SSL labels and synthetic artifacts, which limit what the accuracy numbers mean, but again this is a data-quality limitation rather than a circular derivation. Overall, the evaluation is externally grounded in held-out real data, so no step reduces to its own input by construction.

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

The central claim rests on no new theoretical machinery. The weight sits in five domain assumptions about label fidelity, spectrogram sufficiency, GAN-to-real transfer, segmentation coherence, and GAN stability; five hand-set parameters (filter threshold, segmentation thresholds, augmentation count, noise variance, embedding dimension); and zero invented entities. The augmentation-count entry is internally inconsistent (200 vs 1,000), which affects the augmented training set size directly. The status_SSL label assumption is the heaviest: the authors' own Section 9 states mislabeled samples permeate both classes, so both the baseline and the synthetic-amplified model may be optimizing a noisy target. No parameter in this ledger is fitted to the reported 75% test accuracy; the GAN and CNN hyperparameters come from standard practice.

free parameters (5)
  • cough_detected filter threshold = 0.7
    All samples with cough_detected below 0.7 are dropped (Section 4.2); the threshold is hand-set with no sensitivity analysis and shrinks the already small COVID-19 class.
  • hysteresis segmentation thresholds = high = 2x RMS, low = 0.1x RMS
    Set in Section 4.2; determines how coughs are cut into segments. The authors themselves suspect (Section 8) that the low threshold splits single coughs at the silent compressive phase.
  • synthetic augmentation count per class = 200 (Section 6) or 1,000 (Section 7.2)
    The two stated counts are contradictory; the augmentation volume directly sets the size of the augmented training set used for the 75% result.
  • instance-noise initial variance = 0.1, annealed linearly to 0
    Gaussian noise added to discriminator inputs (Section 6) with a hand-chosen annealing schedule; stability heuristic whose exact schedule affects generated-sample quality.
  • label embedding dimension = 50
    Dimension of the learned class-label embedding in the generator (Section 5.2); chosen without stated justification.
assumptions (5)
  • domain assumption status_SSL labels in the Orlandic (2023) Coughvid revision are an adequate ground truth for healthy vs COVID-19.
    Invoked in Sections 4.2 and 7.1 to define the two classes for both the baseline and the ACGAN, and in Section 7.2 for the augmented classifier. Section 9 concedes substantial mislabeling (self-reported statuses, inconsistent expert review), so if the labels are largely wrong the central 72% to 75% numbers describe artifacts, not disease.
  • domain assumption Mel spectrograms preserve the discriminative cough features used for classification.
    Every model operates on decibel Mel spectrograms (Sections 4.2-4.3). Phase information and fine temporal structure are discarded; the paper presents no evidence that the preserved features carry the COVID-19 signal.
  • domain assumption Synthetic spectrograms from the trained ACGAN are close enough to the real distribution that training on them transfers to real test data.
    This is the mechanism behind the claimed improvement (Sections 6, 7.2). The paper's own Figs. 14-16 show visible artifacts and a dominating discriminator, and the only evidence of transfer is a 3-point single-split gain, so overlap is assumed rather than established.
  • domain assumption The hysteresis segmentation (2x/0.1x RMS thresholds) yields coherent cough units.
    Section 4.2. In Section 8 the authors doubt the threshold separates single coughs into two segments, which would corrupt class-conditional statistics for both the GAN and the classifier.
  • domain assumption Instance noise plus soft labels are sufficient to keep ACGAN training stable and class-conditional.
    Section 6 adopts these heuristics to prevent the discriminator loss from collapsing to zero; the observed loss histories (Fig. 16) show the discriminator still dominates, so the assumption holds only partially.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sound Signal Synthesis with Auxiliary Classifier GAN, COVID-19 cough as an example." pith.science (2026). https://pith.science/paper/2BRE4QIJ

@misc{pith2026250808892,
  author       = {Pith},
  title        = {Pith review of: Sound Signal Synthesis with Auxiliary Classifier GAN, COVID-19 cough as an example},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2BRE4QIJ}},
  note         = {Machine review of arXiv:2508.08892}
}
read the original abstract

One of the fastest-growing domains in AI is healthcare. Given its importance, it has been the interest of many researchers to deploy ML models into the ever-demanding healthcare domain to aid doctors and increase accessibility. Delivering reliable models, however, demands a sizable amount of data, and the recent COVID-19 pandemic served as a reminder of the rampant and scary nature of healthcare that makes training models difficult. To alleviate such scarcity, many published works attempted to synthesize radiological cough data to train better COVID-19 detection models on the respective radiological data. To accommodate the time sensitivity expected during a pandemic, this work focuses on detecting COVID-19 through coughs using synthetic data to improve the accuracy of the classifier. The work begins by training a CNN on a balanced subset of the Coughvid dataset, establishing a baseline classification test accuracy of 72%. The paper demonstrates how an Auxiliary Classification GAN (ACGAN) may be trained to conditionally generate novel synthetic Mel Spectrograms of both healthy and COVID-19 coughs. These coughs are used to augment the training dataset of the CNN classifier, allowing it to reach a new test accuracy of 75%. The work highlights the expected messiness and inconsistency in training and offers insights into detecting and handling such shortcomings.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 32 canonical work pages

  1. [1]

    & Xia, L

    Ai, T., Yang, Z., Hou, H., Zhan, C., Chen, C., Lv, W., ... & Xia, L. (2020). Correlation of chest CT and RT -PCR testing for coronavirus disease 2019 ( COVID-19) in China: a report of 1014 cases. Radiology, 296(2), E32-E40

    work page 2020

  2. [2]

    A., Zhang, N.,

    Bernheim, A., Mei, X., Huang, M., Yang, Y., Fayad, Z. A., Zhang, N., ... & Chung, M. (2020). Chest CT findings in coronavirus disease-19 ( COVID-19): relationship to duration of infection. Radiology, 200463

    work page 2020

  3. [3]

    & Shan, H

    Chung, M., Bernheim, A., Mei, X., Zhang, N., Huang, M., Zeng, X., ... & Shan, H. (2020). CT imaging features of 2019 novel coronavirus (2019- nCoV). Radiology, 295(1), 202-207

    work page 2020

  4. [5]

    Donahue, C., McAuley, J., & Puckette, M. (2018). Adversarial audio synthesis. arXiv preprint arXiv:1802.04208

    arXiv 2018

  5. [6]

    Dravid, A., & Katsaggelos, A. K. (2021). Visual Explanations for Convolutional Neural Networks via Latent Traversal of Generative Adversarial Networks. arXiv e - prints, arXiv-2111

    work page 2021

  6. [7]

    Fang, Y., Zhang, H., Xie, J., Lin, M., Ying, L., Pang, P., & Ji, W. (2020). Sensitivity of chest CT for COVID-19: comparison to RT- PCR. Radiology, 296(2), E115-E117

    work page 2020

  7. [8]

    and Bengio, Y

    Goodfellow, I., Pouget -Abadie, J., Mirza, M., Xu, B., Warde -Farley, D., Ozair, S., Courville, A. and Bengio, Y. (2014). Generative Adversarial Nets. Available at: https://arxiv.org/pdf/1406.2661.pdf

    arXiv 2014

  8. [9]

    J., Ni, Z

    Guan, W. J., Ni, Z. Y., Hu, Y., Liang, W. H., Ou, C. Q., He, J. X., ... & Zhong, N. S. (2020). Clinical characteristics of coronavirus disease 2019 in China. New England journal of medicine, 382(18), 1708- 1720

    work page 2020

Show all 36 references
  1. [10]

    D., Rasmussen, N., Tan, D

    Haritaoglu, E. D., Rasmussen, N., Tan, D. C., Xiao, J., Chaudhari, G., Rajput, A., ... & Pilanci, M. (2022). Using deep learning with large aggregated datasets for COVID- 19 classification from cough. arXiv preprint arXiv:2201.01669

    2022 arXiv

  2. [11]

    Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in neural information processing systems, 33, 6840-6851

    2020

  3. [12]

    & Cao, B

    Huang, C., Wang, Y., Li, X., Ren, L., Zhao, J., Hu, Y., ... & Cao, B. (2020). Clinical features of patients infected with 2019 novel coronavirus in Wuhan, China. The lancet, 395(10223), 497-506

    2020

  4. [13]

    Jiang, Y., Chen, H., Loew, M., & Ko, H. (2020). COVID-19 CT image synthesis with a conditional generative adversarial network. IEEE Journal of Biomedical and Health Informatics, 25(2), 441-452

    2020

  5. [14]

    P., Little, B

    Kanne, J. P., Little, B. P., Chung, J. H., Elicker, B. M., & Ketai, L. H. (2020). Essentials for radiologists on COVID-19: an update—radiology scientific expert panel. Radiology, 296(2), E113-E114

    2020

  6. [15]

    & Comaniciu, D

    Liu, S., Georgescu, B., Xu, Z., Yoo, Y., Chabin, G., Chaganti, S., ... & Comaniciu, D. (2020). 3d tomographic pattern synthesis for enhancing the quantification of COVID-

    2020

  7. [16]

    Mahapatra, D., & Singh, A. (2021). Ct image synthesis using weakly supervised segmentation and geometric inter -label relations for COVID image analysis. arXiv preprint arXiv:2106.10230

    2021 arXiv

  8. [17]

    & Chapman, D

    Menon, S., Mangalagiri, J., Galita, J., Morris, M., Saboury, B., Yesha, Y., ... & Chapman, D. (2021). CCS -GAN: COVID- 19 CT -scan classification with very few positive training images. arXiv preprint arXiv:2110.01605

    2021 arXiv

  9. [18]

    and Backe, B., 2024, September

    Mokayed, H., Alsayed, G., Lodin, F., Hagner, O. and Backe, B., 2024, September. Enhancing object detection in snowy conditions: Evaluating yolo v9 models with augmentation techniques. In 2024 11th International Conference on Internet of Things: Systems, Management and Security...

    2024

  10. [19]

    arXiv preprint arXiv:2005.01903

    2005 arXiv

  11. [20]

    Motamed, S., Rogalla, P., & Khalvati, F. (2021). RANDGAN: randomized generative adversarial network for detection of COVID-19 in chest X -ray. Scientific Reports, 11(1), 1-10

    2021

  12. [21]

    K., Kumar, R.,

    Muguli, A., Pinto, L., Sharma, N., Krishnan, P., Ghosh, P. K., Kumar, R., ... & Nanda, V. (2021). DiCOVA Challenge: Dataset, task, and baseline system for COVID-19 diagnosis using acoustics. arXiv preprint arXiv:2103.09148

    2021 arXiv

  13. [22]

    (2017, July)

    Odena, A., Olah, C., & Shlens, J. (2017, July). Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning (pp. 2642- 2651). PMLR

    2017

  14. [23]

    Orlandic, L., Teijeiro, T., & Atienza, D. (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, 107743

    2023

  15. [24]

    (2015, October)

    Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U -net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer - assisted intervention (pp. 234- 241). Springer, Cham

    2015

  16. [25]

    How GANs assist in COVID- 19 pandemic era: a review

    Saleh, Yahya Sherif Solayman Mohamed, Hamam Mokayed, Konstantina Nikolaidou, Lama Alkhaled, and Yan Chai Hum. "How GANs assist in COVID- 19 pandemic era: a review." Multimedia Tools and Applications 83, no. 10 (2024): 29915- 29944

    2024

  17. [26]

    Sharma, A., & Mishra, P. K. (2022). COVID-MANet: Multi -task attention network for explainable diagnosis and severity assessment of COVID-19 from CXR images. Pattern Recognition, 131, 108826

    2022

  18. [27]

    P., Chhipa, P

    Södergren, I., Nodeh, M. P., Chhipa, P. C., Nikolaidou, K., & Kovács, G. (2021). Detecting COVID-19 from audio recording of coughs using Random Forests and Support Vector Machines. In Interspeech 2021, Brno, Czechia, 30 August -3 September, 2021 (pp. 916- 920). International S...

    2021

  19. [28]

    M., Chowdhury, M

    Tahir, A. M., Chowdhury, M. E., Khandakar, A., Rahman, T., Qiblawey, Y., Khurshid, U., ... & Hamid, T. (2021). COVID-19 infection localization and severity grading from chest X -ray images. Computers in biology and medicine, 139, 105002

    2021

  20. [29]

    Attention is all you need

    Vaswani, A. "Attention is all you need." Advances in Neural Information Processing Systems (2017)

    2017

  21. [30]

    Waheed, A., Goyal, M., Gupta, D., Khanna, A., Al -Turjman, F., & Pinheiro, P. R. (2020). COVIDgan: data augmentation using auxiliary classifier gan for improved COVID-19 detection. Ieee Access, 8, 91916- 91923

    2020

  22. [31]

    Yi, X., Walia, E., & Babyn, P. (2019). Generative adversarial network in medical imaging: A review. Medical Image Analysis, 58, 101552. https://doi.org/10.1016/j.media.2019.101552

    2019

  23. [32]

    Zunair, H., & Hamza, A. B. (2021a). Synthesis of COVID- 19 chest X -rays using unpaired image -toimage translation. Social Network Analysis and Mining, 11(1), 1-12

  24. [33]

    Zunair, H., & Hamza, A. B. (2021b). Synthetic COVID-19 Chest X -ray Dataset for Computer -Aided Diagnosis. arXiv preprint arXiv:2106.09759

    arXiv

  25. [34]

    Zhao, J., Zhang, Y., He, X., & Xie, P. (2020a). COVID-ct-dataset: a ct scan dataset about COVID-19. arXiv preprint arXiv:2003.13865, 490

    2003 arXiv

  26. [35]

    Zhao, W., Zhong, Z., Xie, X., Yu, Q., & Liu, J. (2020b). Relation between chest CT findings and clinical conditions of coronavirus disease ( COVID-19) pneumonia: a multicenter study. AJR Am J Roentgenol, 214(5), 1072-1077

  27. [36]

    Hamdi, S., Oussalah, M., Moussaoui, A., & Saidi, M. (2022). Attention- based hybrid CNN -LSTM and spectral data augmentation for COVID-19 diagnosis from cough sound. Journal of Intelligent Information Systems, 59(2), 367–389

    2022

  28. [37]

    K., Davenport, P

    Lee, K. K., Davenport, P. W., Smith, J. A., Irwin, R. S., McGarvey, L., Mazzone, S. B., Birring, S. S., Abu Dabrh, A., Altman, K. W., Barker, A. F., Birring, S. S., Blackhall, F., Bolser, D. C., Brightling, C., Chang, A. B., Davenport, P., El Solh, A. A., Escalante, P., Field,...

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.