Adaptive Differential Denoising for Respiratory Sounds Classification

Gaoyang Dong; Minghui Zhang; Ping Sun; Zhicheng Zhang

REVIEW 2 major objections 6 minor 38 references

Adaptive Differential Denoising for Respiratory Sounds Classification

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that building denoising directly into a respiratory sound classifier, through a learnable spectral filter, differential attention, and a label-smoothed loss, pushes the ICBHI 2017 benchmark score to 65.53%, exceeding the…

desk verdict Plausible 2-point SOTA gain on ICBHI 2017, but the missing plain-AST control and a self-contradicted novelty claim keep it from being fully supported as written. read the letter →

arxiv 2506.02505 v1 pith:KA5ZJRXY submitted 2025-06-03 eess.AS cs.SD

Gaoyang Dong , Zhicheng Zhang , Ping Sun , Minghui Zhang This is my paper

classification eess.AScs.SD

keywords respiratorysoundclassificationadaptivedifferentialdenoisingfrequencyfilterattentionlabelsmoothingICBHI2017audiospectrogramtransformer

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

Respiratory sound classification is hampered by noise that overlaps with diagnostic features, and previous methods either denoise separately or use generic audio denoisers that can discard pathological content. This paper tries to establish that denoising can be learned implicitly inside the classifier, using only categorical labels, by combining a learnable spectral filter, a differential attention layer, and a label-smoothed loss. On the ICBHI 2017 benchmark, the proposed Adaptive Differential Denoising network reports a Score of 65.53%, which the authors say surpasses the prior best method by 1.99%. If the result holds, it would make robust respiratory sound classification achievable in real clinical recordings without paired clean/noisy data.

What carries the argument

The argument rests on three jointly trained components. The Adaptive Frequency Filter (AFF) moves the Mel spectrogram into the frequency domain with the FFT, multiplies it by a learnable instance-adaptive mask, applies soft shrinkage $S_\alpha(x)=\operatorname{sign}(x)\max\{|x|-\alpha,0\}$, and transforms back, suppressing noise while preserving sparse high-frequency content. The Differential Denoise Layer (DDL) computes multi-head differential attention: it projects the input into two query/key pairs, applies two softmax attention maps, and subtracts one from the other (scaled by a learnable $\lambda$) before attending to the values, which cancels noise-sensitive variations and keeps stable features. The hybrid loss $\mathcal{L}=\beta\,\mathcal{L}_{\text{BiasDenoise}}+(1-\beta)\mathcal{L}_{\text{CE}}$ combines a label-smoothed cross-entropy term—used as a denoising guide without clean references—with standard classification, with $\beta=0.5$.

What would settle it

If one re-runs the same AST backbone and preprocessing on ICBHI 2017 with all three proposed modules disabled and still obtains a Score above 65.53% (or within 1.99% of the reported value), the attribution of the gain to these modules collapses. Alternatively, a sensitivity sweep over $\alpha\in\{0.005,0.02,0.1\}$, $\beta\in\{0,0.5,1\}$, and $\epsilon\in\{0,0.2,0.5\}$ that shows Score swings greater than 1.99% would indicate the headline number is not stable.

Watch

Extended reading notes

Core claim

The paper's central claim is that its Adaptive Differential Denoising (ADD) network, which couples three modules—an adaptive frequency filter, a differential denoise layer, and a bias denoising loss—improves respiratory sound classification on the ICBHI 2017 benchmark to a Score of 65.53%, surpassing the previous best method (BTS) by 1.99%. The authors state that this is the first deep-learning denoising approach designed specifically for respiratory sounds, and that it works without paired clean/noisy training data because the loss only needs categorical labels. Ablation results show each module contributes, with the full combination scoring higher than any pair.

Load-bearing premise

The reported 1.99% improvement over the prior state of the art is attributed to the three proposed modules rather than to unstated differences in preprocessing, backbone initialization, or hyperparameter choices; in particular, the hyperparameters $\alpha=0.02$, $\beta=0.5$, $\epsilon=0.2$, and the 8-second clip length are fixed without sensitivity analysis.

Editorial extensions

If this is right

Respiratory sound classifiers could be deployed in noisier clinical environments without a separate denoising stage, since the model learns to suppress noise and preserve pathology from labels alone.
The same three modules can be inserted into other audio backbones, because the paper shows gains on both a CNN (ResNet50) and a Transformer (AST) architecture.
Improving both sensitivity and specificity together, rather than trading one off, suggests the model is separating signal from noise rather than tilting the decision boundary.
The implicit-denoising objective offers a template for other physiological acoustic signals, such as heart or bowel sounds, where clean reference recordings are equally unavailable.

Reading between the lines

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

The fixed choice of 8-second clips and cyclic padding is not ablated, so a controlled re-run of all baselines under identical preprocessing would test whether the reported 1.99% margin survives.
The label-smoothing parameter $\epsilon$ is described as an uncertainty buffer; a testable extension is to vary $\epsilon$ during inference to adjust the specificity/sensitivity balance without retraining, which the paper does not examine.
Because differential attention originates in text models, its effectiveness here suggests that comparing augmented views of the input could help denoise other noisy bio-signals, an extension the paper leaves implicit.

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

Quick take: this is an honest empirical architecture paper with an internally consistent ablation, but the central SOTA claim is under-supported because there is no plain-AST baseline, and the "first to explore DL denoising for respiratory sounds" claim is contradicted by their own reference [18].

What's actually new: the specific combination of an adaptive frequency filter (from Huang et al.), differential attention (from Ye et al.'s Differential Transformer), and a label-smoothed denoising loss, evaluated on ICBHI 2017. The reported 65.53% Score is higher than anything they cite. The ablation (Table 2) shows each component adding a bit, from 63.47 to 65.53, and the method is described clearly enough to reproduce. The code link helps. That is real, if incremental, progress in a small subfield.

Soft spots: the main problem is the missing baseline. Table 2's first row already includes AFF, so there is no row for plain AST with the same preprocessing, clip length, optimizer, and epochs. The 1.99% claimed gain over BTS (Table 1) could easily come from their 8-second truncation/padding, 64 Mel filters, 16 kHz resampling, label smoothing, or training choices, rather than from the three modules. The comparison table relies on numbers quoted from PapersWithCode [24], which likely involve different setups. This is an omitted control, not an internal inconsistency, but it is exactly the control needed to support the SOTA statement. Second, there are no error bars or multiple-seed results; with a 1.99% gap and single-run numbers, the margin could be noise. Third, the Introduction's claim to be "the first to explore deep learning-based denoising techniques specifically for respiratory sounds" is directly contradicted by reference [18], Tzeng et al. 2024, which does DL-based audio enhancement for RSC. That overclaim should be corrected, though it does not affect the measured performance.

Who this is for: anyone working on respiratory sound classification or using AST on medical audio will want to know this result and will want to see the missing control added. The paper deserves a serious referee, but it needs revision before acceptance: add a plain-AST baseline, report variance across seeds, do a small sensitivity check on alpha/beta/epsilon, and fix the novelty statement.

Referee Report

2 major / 6 minor

Summary. The paper proposes an Adaptive Differential Denoising (ADD) network for respiratory sound classification, combining an Adaptive Frequency Filter (AFF), a Differential Denoise Layer (DDL), and a bias-aware denoising loss. The method is evaluated on the ICBHI 2017 benchmark, where the AST-based ADD model reports a Score of 65.53%, surpassing the prior best BTS (63.54%) by 1.99%. The authors claim three contributions: an implicit denoising approach that does not require clean reference signals, the ADD architecture itself, and state-of-the-art performance on ICBHI 2017. Code is made publicly available.

Significance. If the performance claim is robust, the architecture offers a practical way to integrate denoising directly into the classification pipeline without paired clean/noisy data, which is a relevant engineering contribution. The internal ablation is monotone and consistent with the proposed design, and the release of code supports reproducibility. However, the lack of a controlled same-protocol baseline and the absence of variance reporting mean that the 1.99% margin over BTS is not yet fully attributable to the proposed modules. The work is of moderate significance for the respiratory sound classification community, but the central SOTA claim requires stronger empirical support before it can be accepted as stated.

major comments (2)

[§3.2.2, Table 2] The ablation study does not include a plain-AST baseline; the first row already contains AFF ('w AFF', 63.47). Because Table 1 compares the full model against BTS and other published methods, the reported 1.99% improvement over BTS cannot be isolated to the ADD modules. The improvement could stem entirely from the 8-second clip length, the cyclic padding/truncation, the Mel-spectrogram preprocessing, the label-smoothed loss term, or other training choices that differ from the BTS setup. I request a same-protocol baseline obtained by fine-tuning the same AST backbone with identical preprocessing and hyperparameters but without AFF, DDL, and the bias denoising loss; ideally, BTS should also be re-run under the same protocol.
[§3.1.3 and §3.2.2] All fixed hyperparameters (alpha=0.02, beta=0.5, epsilon=0.2), the choice of the 8-second input length, and the architecture configuration are set to single values with no sensitivity analysis or multiple-seed results. The claimed margin over BTS is only 1.99 percentage points in Score, which is an average of sensitivity and specificity; without error bars or confidence intervals, the reported difference may be within run-to-run or test-set noise. I request means and standard deviations over at least three random seeds (or an equivalent variance estimate) and, if space permits, a small sensitivity analysis for alpha, beta, and epsilon.

minor comments (6)

[§1, contribution bullet 1] The statement that this work is 'the first to explore deep learning-based denoising techniques specifically for respiratory sounds' is contradicted by the paper's own reference [18], which presents deep learning-based audio enhancement for respiratory sound classification; I recommend softening or more carefully qualifying this claim.
[§2.1] The sentence 'We determined that an 8-second length yields optimal performance' is not supported by an experiment in the paper; please provide the evidence or rephrase it as a design choice.
[§2.2, Eq. (1)] The notation uses X(t, f) for the input spectrogram and X(U, V) for its modulation spectrum, which may confuse readers; using a different symbol for the transformed quantity would improve clarity.
[Table 1 caption] The caption says the results are 'from [24]' (PapersWithCode), but individual rows also cite specific papers; please clarify which numbers are quoted from the public benchmark and which are taken directly from the original publications.
[§2.4, Eq. (6)] The loss term is written with a space as 'L_Bias Denoise' while the text uses 'L_BiasDenoise'; please make the notation consistent.
[General] The manuscript alternates between 'SOTA' and 'sota' in the abstract and introduction; please unify the capitalization.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the reported gain is an empirical benchmark comparison, not a derivation from fitted inputs; minor self-citations are not load-bearing.

full rationale

This is an empirical architecture paper; there is no claimed derivation chain that reduces to its own inputs. The three proposed components (AFF, DDL, bias denoising loss) are trainable modules or loss terms optimized with the standard cross-entropy objective on the ICBHI 2017 training split, and the reported 65.53% Score is measured on the official test split. Equation (6) combines the classification loss with a label-smoothed cross-entropy term (Equation 7); although the name 'bias denoising loss' is a new label for a known regularization technique, the term is not fitted to or derived from the test metric, so no self-definitional or fitted-input-as-prediction circularity arises. The authors' self-citations [10] and [12] describe prior respiratory sound classification work and are not used to justify the central claim. The table of state-of-the-art numbers is quoted from an external PapersWithCode collection [24], and the comparison is therefore a benchmark comparison rather than an internally forced result. Two non-circular weaknesses should be noted: the ablation (Table 2) lacks a plain-AST baseline trained under identical preprocessing and hyperparameters, so the 1.99% margin over BTS is not fully controlled, and the Introduction's claim to be 'the first to explore deep learning-based denoising techniques specifically for respiratory sounds' is contradicted by the paper's own reference [18], a deep-learning audio-enhancement study for respiratory sound classification. These are correctness and novelty concerns, not circularity.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method does not introduce new physical entities. Its 'Adaptive Frequency Filter' and 'Differential Denoise Layer' are re-brandings of published components (AFF from [21], MHDA from [22]). The free parameters listed are hand-chosen hyperparameters that are not supported by sensitivity experiments, which is the main burden the central claim carries.

free parameters (5)

alpha (soft shrink threshold) = 0.02
Set in Section 3.1.3 to control sparsity in the adaptive frequency filter. No sensitivity analysis is reported.
beta (loss weight) = 0.5
Balances the bias denoising loss and cross-entropy in Eq. 6. Chosen without ablation or sensitivity study.
epsilon (label smoothing) = 0.2
Used in Eq. 7 for the 'bias denoising' loss. Fixed in Section 3.1.3 with no reported sensitivity analysis.
input clip length = 8 seconds
The paper states 'We determined that an 8-second length yields optimal performance' (Section 2.1) but gives no supporting experiments.
lambda (differential attention scale) = learnable, not reported
Learnable scaling factor in Eq. 5; its final trained value is not given, so its behavior is unverifiable.

assumptions (3)

domain assumption The ICBHI 2017 official train/test split and evaluation protocol are correctly applied.
The paper states it follows the official protocol (Section 3.1.1), but no code or split files are provided to verify no patient overlap or exact preprocessing.
standard math The Fourier transform and convolution theorem hold exactly for the discrete spectrogram as used in Eqs. 1-2.
Invoked in Section 2.2 without proof, standard background.
domain assumption Pretrained AudioSet features (ResNet50, AST-Base) transfer positively to respiratory sound classification.
The backbones are initialized from AudioSet weights; the paper does not test randomly initialized versions, so the SOTA claim assumes this transfer is beneficial.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Differential Denoising for Respiratory Sounds Classification." pith.science (2026). https://pith.science/paper/KA5ZJRXY

@misc{pith2026250602505,
  author       = {Pith},
  title        = {Pith review of: Adaptive Differential Denoising for Respiratory Sounds Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KA5ZJRXY}},
  note         = {Machine review of arXiv:2506.02505}
}

read the original abstract

Automated respiratory sound classification faces practical challenges from background noise and insufficient denoising in existing systems. We propose Adaptive Differential Denoising network, that integrates noise suppression and pathological feature preservation via three innovations: 1) Adaptive Frequency Filter with learnable spectral masks and soft shrink to eliminate noise while retaining diagnostic high-frequency components; 2) A Differential Denoise Layer using differential attention to reduce noise-induced variations through augmented sample comparisons; 3) A bias denoising loss jointly optimizing classification and robustness without clean labels. Experiments on the ICBHI2017 dataset show that our method achieves 65.53\% of the Score, which is improved by 1.99\% over the previous sota method. The code is available in https://github.com/deegy666/ADD-RSC

Figures

Figures reproduced from arXiv: 2506.02505 by the authors.

**Figure 1.** Overall illustration of proposed model architecture. (a) shows Adaptive Differential Denoising Framework and (b) shows the specific details of the Multi-Head Differential Attention (MHDA) Mechanism. where ϕ is the network for mapping the previously denoised features, specifically we use a 1 × 1 convolutional layer to represent it, yc is the true label, C is the total number of classifications, and p is the feature… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 28 canonical work pages

[18]

Speech enhancement using spec- tral subtraction-type algorithms: A comparison and simulation study,

N. Upadhyay and A. Karmakar, “Speech enhancement using spec- tral subtraction-type algorithms: A comparison and simulation study,”Procedia Computer Science, vol. 54, pp. 574–584, 2015

work page 2015
[24]

Mossformer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,

S. Zhao and B. Ma, “Mossformer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

work page 2023
[1]

Introduction Automated respiratory sound classification (ARSC) has emerged as a critical tool for the non-invasive diagnosis of pul- monary diseases such as asthma, pneumonia, and chronic ob- structive pulmonary disease (COPD) [1]. By analyzing aus- cultation signals captured via digital stethoscopes or wearable sensors, ARSC systems aim to categorize pat...

work page Pith review arXiv 2000
[2]

Proposed Methods In this section, we introduce the Adaptive Differential Denois- ing framework for RSC. The overall pipeline is illustrated in Figure 1, which consists of three key components: (1) Adaptive Frequency Filters (AFF), (2) Differential Denoise Layer (DDL), and (3) Loss Functions. Below, we describe each component in detail. 2.1. Preprocessing ...

work page
[3]

Experimental Setup 3.1.1

Experiments 3.1. Experimental Setup 3.1.1. Dataset This experiment employs the ICBHI 2017 dataset [23], the largest publicly available respiratory sound corpus, for evalu- ation. Audio data were collected from 126 participants. Due to the complexity of the clinical recording environment, the original recordings may contain a variety of real noise distur- ...

work page 2017
[4]

We introduce a novel end-to-end implicit denois- ing approach for respiratory sound classification, designed for datasets lacking dedicated denoising procedures

Conclusion In this work, we alleviate the existing gap in respiratory sound denoising. We introduce a novel end-to-end implicit denois- ing approach for respiratory sound classification, designed for datasets lacking dedicated denoising procedures. The method initially suppresses the majority of noise via the AFF tech- nique, followed by a fine-grained no...

work page
[5]

82260024

Acknowledgements This research is supported by the National Natural Science Foundation of China under Grant No. 82260024. Acknowl- edgments to the First Affiliated Hospital of Nanchang Univer- sity and the Laboratory of Imaging and Visual Representation of Nanchang University

work page
[6]

Au- tomatic adventitious respiratory sound analysis: A systematic re- view,

R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, “Au- tomatic adventitious respiratory sound analysis: A systematic re- view,”PloS one, vol. 12, no. 5, p. e0177926, 2017

work page 2017

Show all 38 references

[7]

Telemedicine for healthcare: Capabilities, features, barriers, and applications,

A. Haleem, M. Javaid, R. P. Singh, and R. Suman, “Telemedicine for healthcare: Capabilities, features, barriers, and applications,” Sensors international, vol. 2, p. 100117, 2021

2021
[8]

Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,

S. Gairola, F. Tom, N. Kwatra, and M. Jain, “Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,” in2021 43rd Annual International Confer- ence of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2021, pp. 527–530

2021
[9]

Repaugment: Input-agnostic representation-level aug- mentation for respiratory sound classification,

J.-W. Kim, M. Toikkanen, S. Bae, M. Kim, and H.-Y . Jung, “Repaugment: Input-agnostic representation-level aug- mentation for respiratory sound classification,”arXiv preprint arXiv:2405.02996, 2024

2024 arXiv
[10]

Ad- versarial Fine-tuning using Generated Respiratory Sound to ad- dress class imbalance,

J.-W. Kim, C. Yoon, M. Toikkanen, S. Bae, and H.-Y . Jung, “Ad- versarial Fine-tuning using Generated Respiratory Sound to ad- dress class imbalance,”arXiv preprint arXiv:2311.06480, 2023

2023 arXiv
[11]

Example-based explanations with adversarial attacks for respira- tory sound analysis,

Y . Chang, Z. Ren, T. T. Nguyen, W. Nejdl, and B. W. Schuller, “Example-based explanations with adversarial attacks for respira- tory sound analysis,”arXiv preprint arXiv:2203.16141, 2022

2022 arXiv
[12]

A domain transfer based data augmenta- tion method for automated respiratory classification,

Z. Wang and Z. Wang, “A domain transfer based data augmenta- tion method for automated respiratory classification,” inICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022, pp. 9017–9021

2022
[13]

Stethoscope-guided Supervised Contrastive Learning for Cross- domain Adaptation on Respiratory Sound Classification,

J.-W. Kim, S. Bae, W.-Y . Cho, B. Lee, and H.-Y . Jung, “Stethoscope-guided Supervised Contrastive Learning for Cross- domain Adaptation on Respiratory Sound Classification,”arXiv preprint arXiv:2312.09603, 2023

2023 arXiv
[14]

Patch-Mix Contrastive Learn- ing with Audio Spectrogram Transformer on Respiratory Sound Classification,

S. Bae, J.-W. Kim, W.-Y . Cho, H. Baek, S. Son, B. Lee, C. Ha, K. Tae, S. Kim, and S.-Y . Yun, “Patch-Mix Contrastive Learn- ing with Audio Spectrogram Transformer on Respiratory Sound Classification,”arXiv preprint arXiv:2305.14032, 2023

2023 arXiv
[15]

Respiratory sounds classification by fusing the time-domain and 2d spectral features,

G. Dong, Y . Shen, J. Wang, M. Zhang, P. Sun, and M. Zhang, “Respiratory sounds classification by fusing the time-domain and 2d spectral features,”Biomedical Signal Processing and Control, vol. 107, p. 107790, 2025

2025
[16]

Respi- ratory Sound Classification Based on Swin Transformer,

W. Sun, F. Zhang, P. Sun, Q. Hu, J. Wang, and M. Zhang, “Respi- ratory Sound Classification Based on Swin Transformer,” in2023 8th International Conference on Signal and Image Processing (ICSIP). IEEE, 2023, pp. 511–515

2023
[17]

Lightweight hierarchical transformer combining patch-random and positional encoding for respiratory sound classification,

J. Wang, G. Dong, Y . Shen, M. Zhang, and P. Sun, “Lightweight hierarchical transformer combining patch-random and positional encoding for respiratory sound classification,” in2024 9th Inter- national Conference on Signal and Image Processing (ICSIP). IEEE, 2024, pp. 580–584

2024
[19]

Speech enhancement based on hypothesized wiener filtering

V . Ramasubramanian and D. Vijaywargi, “Speech enhancement based on hypothesized wiener filtering.” inINTERSPEECH, 2008, pp. 167–170

2008
[20]

Design and implementation of butterworth, chebyshev-i and elliptic filter for speech signal analysis,

P. Podder, M. M. Hasan, M. R. Islam, and M. Sayeed, “Design and implementation of butterworth, chebyshev-i and elliptic filter for speech signal analysis,”arXiv preprint arXiv:2002.03130, 2020

2002 arXiv
[21]

An improved method of audio denoising based on wavelet transform,

M. Ali and P. Shemi, “An improved method of audio denoising based on wavelet transform,” in2015 international conference on Power, Instrumentation, Control and Computing (PICC). IEEE, 2015, pp. 1–6

2015
[22]

The physiological basis of high-frequency oscillatory ventilation and current evidence in adults and children: a narrative review,

A. G. Miller, H. L. Tan, B. J. Smith, A. T. Rotta, and J. H. Lee, “The physiological basis of high-frequency oscillatory ventilation and current evidence in adults and children: a narrative review,” Frontiers in physiology, vol. 13, p. 813478, 2022

2022
[23]

Tzeng, J.-L

J.-T. Tzeng, J.-L. Li, H.-Y . Chen, C.-H. Huang, C.-H. Chen, C.-Y . Fan, E. P.-C. Huang, and C.-C. Lee, “Improving robustness and clinical applicability of automatic respiratory sound classification using deep learning-based audio enhancement: Algorithm devel- opment and valid...

2024 arXiv
[25]

Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,

Y . Luo and N. Mesgarani, “Conv-tasnet: Surpassing ideal time– frequency magnitude masking for speech separation,”IEEE/ACM transactions on audio, speech, and language processing, vol. 27, no. 8, pp. 1256–1266, 2019

2019
[26]

Adap- tive frequency filters as efficient global token mixers,

Z. Huang, Z. Zhang, C. Lan, Z.-J. Zha, Y . Lu, and B. Guo, “Adap- tive frequency filters as efficient global token mixers,” inProceed- ings of the IEEE/CVF International Conference on Computer Vi- sion, 2023, pp. 6049–6059

2023
[27]

Differential transformer,

T. Ye, L. Dong, Y . Xia, Y . Sun, Y . Zhu, G. Huang, and F. Wei, “Differential transformer,”arXiv preprint arXiv:2410.05258, 2024

2024 arXiv
[28]

A respiratory sound database for the development of auto- mated classification,

B. Rocha, D. Filos, L. Mendes, I. V ogiatzis, E. Perantoni, E. Kaimakamis, P. Natsiavas, A. Oliveira, C. J´acome, A. Marques et al., “A respiratory sound database for the development of auto- mated classification,” inPrecision Medicine Powered by pHealth and Connected Health: ...

2017
[29]

Icbhi respiratory sound database: Audio classification state- of-the-art on papers with code,

“Icbhi respiratory sound database: Audio classification state- of-the-art on papers with code,” 2025, accessed: 2025- 02-18. [Online]. Available: https://paperswithcode.com/sota/ audio-classification-on-icbhi-respiratory

2025
[30]

Lungbrn: A smart digital stethoscope for detecting respiratory disease using bi-resnet deep learning algorithm,

Y . Ma, X. Xu, Q. Yu, Y . Zhang, Y . Li, J. Zhao, and G. Wang, “Lungbrn: A smart digital stethoscope for detecting respiratory disease using bi-resnet deep learning algorithm,” in2019 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2019, pp. 1–4

2019
[31]

Lung sound classification using co-tuning and stochastic normalization,

T. Nguyen and F. Pernkopf, “Lung sound classification using co-tuning and stochastic normalization,”IEEE Transactions on Biomedical Engineering, vol. 69, no. 9, pp. 2872–2882, 2022

2022
[32]

Example-based Explanations with Adversarial Attacks for Res- piratory Sound Analysis,

Y . Chang, Z. Ren, T. T. Nguyen, W. Nejdl, and B. W. Schuller, “Example-based Explanations with Adversarial Attacks for Res- piratory Sound Analysis,” inProc. Interspeech 2022, 2022, pp. 4003–4007

2022
[33]

Supervised contrastive learning for respiratory sound classification,

I. Moummad and N. Farrugia, “Supervised contrastive learning for respiratory sound classification,”arXiv preprint arXiv:2210.16192, 2022

2022 arXiv
[34]

Masked modeling duo: Towards a universal audio pre-training framework,

D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked modeling duo: Towards a universal audio pre-training framework,”IEEE/ACM Transactions on Audio, Speech, and Lan- guage Processing, 2024

2024
[35]

Bts: Bridging text and sound modalities for metadata-aided res- piratory sound classification,

J.-W. Kim, M. Toikkanen, Y . Choi, S.-E. Moon, and H.-Y . Jung, “Bts: Bridging text and sound modalities for metadata-aided res- piratory sound classification,” inInterspeech 2024, 2024, pp. 1690–1694

2024
[36]

Multi-View Spec- trogram Transformer for Respiratory Sound Classification,

W. He, Y . Yan, J. Ren, R. Bai, and X. Jiang, “Multi-View Spec- trogram Transformer for Respiratory Sound Classification,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 8626– 8630

2024
[37]

Lungadapter: Efficient adapting audio spectrogram transformer for lung sound classifica- tion,

L. Xiao, L. Fang, Y . Yang, and W. Tu, “Lungadapter: Efficient adapting audio spectrogram transformer for lung sound classifica- tion,” inProc. Interspeech 2024, 2024, pp. 4738–4742

2024
[38]

Cycle- guardian: A framework for automatic respiratorysound classifica- tion based on improved deep clustering and contrastive learning,

Y . Chu, Q. Wang, E. Zhou, L. Fu, Q. Liu, and G. Zheng, “Cycle- guardian: A framework for automatic respiratorysound classifica- tion based on improved deep clustering and contrastive learning,” arXiv preprint arXiv:2502.00734, 2025

2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [18]

Speech enhancement using spec- tral subtraction-type algorithms: A comparison and simulation study,

N. Upadhyay and A. Karmakar, “Speech enhancement using spec- tral subtraction-type algorithms: A comparison and simulation study,”Procedia Computer Science, vol. 54, pp. 574–584, 2015

work page 2015

[2] [24]

Mossformer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,

S. Zhao and B. Ma, “Mossformer: Pushing the performance limit of monaural speech separation using gated single-head transformer with convolution-augmented joint self-attentions,” in ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023, pp. 1–5

work page 2023

[3] [1]

Introduction Automated respiratory sound classification (ARSC) has emerged as a critical tool for the non-invasive diagnosis of pul- monary diseases such as asthma, pneumonia, and chronic ob- structive pulmonary disease (COPD) [1]. By analyzing aus- cultation signals captured via digital stethoscopes or wearable sensors, ARSC systems aim to categorize pat...

work page Pith review arXiv 2000

[4] [2]

Proposed Methods In this section, we introduce the Adaptive Differential Denois- ing framework for RSC. The overall pipeline is illustrated in Figure 1, which consists of three key components: (1) Adaptive Frequency Filters (AFF), (2) Differential Denoise Layer (DDL), and (3) Loss Functions. Below, we describe each component in detail. 2.1. Preprocessing ...

work page

[5] [3]

Experimental Setup 3.1.1

Experiments 3.1. Experimental Setup 3.1.1. Dataset This experiment employs the ICBHI 2017 dataset [23], the largest publicly available respiratory sound corpus, for evalu- ation. Audio data were collected from 126 participants. Due to the complexity of the clinical recording environment, the original recordings may contain a variety of real noise distur- ...

work page 2017

[6] [4]

We introduce a novel end-to-end implicit denois- ing approach for respiratory sound classification, designed for datasets lacking dedicated denoising procedures

Conclusion In this work, we alleviate the existing gap in respiratory sound denoising. We introduce a novel end-to-end implicit denois- ing approach for respiratory sound classification, designed for datasets lacking dedicated denoising procedures. The method initially suppresses the majority of noise via the AFF tech- nique, followed by a fine-grained no...

work page

[7] [5]

82260024

Acknowledgements This research is supported by the National Natural Science Foundation of China under Grant No. 82260024. Acknowl- edgments to the First Affiliated Hospital of Nanchang Univer- sity and the Laboratory of Imaging and Visual Representation of Nanchang University

work page

[8] [6]

Au- tomatic adventitious respiratory sound analysis: A systematic re- view,

R. X. A. Pramono, S. Bowyer, and E. Rodriguez-Villegas, “Au- tomatic adventitious respiratory sound analysis: A systematic re- view,”PloS one, vol. 12, no. 5, p. e0177926, 2017

work page 2017