Pith. sign in

REVIEW 5 major objections 6 minor 30 references

Waveform-Logmel Audio Neural Networks for Respiratory Sound Classification

T0 review · 5 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read This paper claims that a dual-input network reading both raw waveform and log-mel spectrogram, fused and context-modeled by a bidirectional GRU, achieves top published results on the SPRSound pediatric respiratory dataset.

desk verdict A competent architecture-plus-ablation paper whose SOTA claim is not verifiable as written: the test split is unnamed and the binary metrics can be carried by just two frequent classes. read the letter →

arxiv 2504.17156 v1 pith:PEBEY2KE submitted 2025-04-24 cs.SD

classification cs.SD
keywords respiratorysoundclassificationwaveform-logmelaudioneuralnetworksSpectrogramTransformerBi-GRUmulti-classfocallossSPRSounddatasetcomputer-assistedauscultation
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 sounds are hard to classify automatically because abnormal events are scarce and because a single input representation loses information: waveforms lack frequency detail, and log-mel spectrograms lose fine structure at high frequencies and have low temporal resolution. This paper proposes WLANN, a network that reads both the raw waveform and the log-mel spectrogram, fuses the two feature streams, and uses a bidirectional GRU to model how frames evolve over time. On the SPRSound pediatric dataset the model reaches 90.3% sensitivity and a 93.6% total score, which the authors report as the best among the published methods they compare with. If the result holds, automated auscultation moves closer to being a usable clinical aid for detecting abnormal breath sounds like crackles and wheezes.

What carries the argument

The load-bearing object is the fusion tensor $C \in \mathbb{R}^{F \times T \times (C+C/F)}$: the waveform branch's 1D-CNN output is reshaped into $F$ frequency bins and concatenated with the AST branch's $C$-channel spectrogram output along the channel axis. This fused frequency–time representation is mean-pooled over frequency, then modeled by a bidirectional gated recurrent unit (Bi-GRU), a network that reads the frame sequence forward and backward; a linear layer with sigmoid activation maps to class labels. Multi-class focal loss with $\gamma=2$ down-weights the abundant normal class so training focuses on abnormal sounds. The argument is carried by this particular fusion: waveform features supply time-domain detail the spectrogram loses, the spectrogram supplies frequency structure the waveform lacks, and Bi-GRU supplies the frame-to-frame context that makes abnormal events recognizable.

What would settle it

Run WLANN under the paper's training recipe on the SPRSound inter-patient split and report per-class sensitivity for Rhonchi, Stridor, Coarse Crackle, Fine Crackle, and Wheeze; the central claim would fail if the inter-patient total score falls below the 90.9% DenseNet169 baseline in the paper's Table II, or if rare classes show near-zero sensitivity.

Watch

Extended reading notes

Core claim

The central claim is that a dual-branch architecture — a 1D-CNN reading the raw waveform and an Audio Spectrogram Transformer reading the log-mel spectrogram — whose outputs are concatenated along the channel axis and passed through a Bi-GRU, classifies SPRSound respiratory events better than any previously published system the authors compare against. The reported numbers are 90.3% sensitivity, 96.9% specificity, and 93.6% total score, versus 90.9% total score for the strongest prior entry in the comparison table. Ablation results in the paper show that the fused input outperforms either branch alone, and that adding Bi-GRU raises sensitivity from 86.1% to 90.3%, supporting the paper's conclusion that frame-level temporal context is what makes abnormal sounds detectable.

Load-bearing premise

The headline numbers assume that all methods in the comparison table were evaluated on the same SPRSound test split with the same preprocessing, but the paper does not state which split produced its 90.3% sensitivity and 93.6% total score.

Editorial extensions

If this is right

  • A model with the reported sensitivity and specificity balance — 90.3% sensitivity at 96.9% specificity — would detect abnormal respiratory events in about nine out of ten cases while keeping false alarms low.
  • The ablation table shows that the full WLANN beats both single-branch models: fused waveform plus spectrogram without Bi-GRU scores 91.2% total score, and adding Bi-GRU raises it to 93.6%.
  • Multi-class focal loss with $\gamma=2$ is part of the recipe; it down-weights normal sounds so that training concentrates on the scarce abnormal classes.
  • The architecture is assembled from publicly available components (AST, 1D-CNN, Bi-GRU) and evaluated on an open dataset, so the comparison can be replicated and extended.
  • If the result holds, automated stethoscope analysis can move from simply identifying normal breathing to reliably flagging crackles and wheezes, the step needed for clinical use.

Reading between the lines

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

  • The paper reports only pooled abnormal-event sensitivity; a per-class breakdown would clarify whether the 90.3% reflects genuine detection of rare classes such as Stridor and Rhonchi or is dominated by the abundant Fine Crackle and Wheeze events.
  • Because the 40–850 Hz Butterworth filter removes content above 850 Hz, the waveform branch may be blind to high-frequency wheeze harmonics; testing a wider passband could improve or clarify the fusion's contribution.
  • The same waveform-plus-spectrogram-plus-Bi-GRU recipe is domain-agnostic and could transfer to other auscultation signals such as heart or bowel sounds, where abnormal events are also rare and class-imbalanced.
  • If the reported numbers come from the intra-patient split, the clinically relevant next result is inter-patient performance; that number would decide whether the gain generalizes to patients not seen in training.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper proposes WLANN, a neural architecture for respiratory sound classification that fuses features from a waveform-processing 1D-CNN and an Audio Spectrogram Transformer (AST), then applies a Bi-GRU for frame-level context modeling. The model is trained with multi-class focal loss and evaluated on the SPRSound pediatric respiratory dataset. The authors report 90.3% sensitivity and 93.6% total score, claiming state-of-the-art performance and the ability to distinguish pathological respiratory sound classes. An ablation study in Table III shows that each architectural component contributes to the reported performance.

Significance. If the reported results are reproducible and the comparison is fair, WLANN would be a meaningful contribution to respiratory sound classification: the fusion of waveform and spectrogram features with recurrent context modeling is a sensible design, and the ablation in Table III credibly attributes gains to each component. The use of the public SPRSound dataset and the comparison with several published methods are also strengths. However, the significance is currently limited by unverifiable protocol details: it is not stated which test split produced the headline numbers, the metrics pool all abnormal classes into a single binary positive class, and no per-class results are reported, so the central claims of state-of-the-art performance and multiclass discrimination are not yet established.

major comments (5)
  1. [Section IV-D, Table II] The manuscript does not state which SPRSound test split produced the results in Table II. Table I defines two official splits, Testing-1 (intra-patient) and Testing-2 (inter-patient), and the authors themselves describe Testing-2 as the more robust generalization split. Because the two splits have very different patient overlap and class distributions, the headline values of 90.3% sensitivity and 93.6% total score are not interpretable without specifying the split, and the comparison with TRespNET and other cited methods is not a controlled comparison unless all methods were evaluated on the same split with the same preprocessing. Please state the split for every table and report results on both splits.
  2. [Section IV-C, Eqs. (2)-(6)] The evaluation metrics micro-average over pooled abnormal events, so a model that detects only the frequent classes could achieve high sensitivity while failing on rare pathological classes. From Table I, Wheeze and Fine Crackle together constitute 99% of abnormal events in Testing-2 (305+80 of 389) and 89.6% in Testing-1 (108+175 of 316). The abstract and conclusion claim that the model distinguishes pathological respiratory sound classes, but the paper reports no per-class sensitivity, specificity, or confusion matrix. The multiclass discrimination claim is therefore not supported by the provided evidence. Please add per-class results and, if appropriate, macro-averaged metrics.
  3. [Section IV-D, Table II] The comparison with previous methods lacks essential protocol details. It is not stated whether the cited numbers in Table II are taken from the original publications, re-implemented by the authors, or evaluated on the same test split with the same event-level pooling and the same preprocessing chain. Without this information, the state-of-the-art claim is not verifiable. Please specify the origin of each compared result and provide a reproducible evaluation protocol, ideally with code or a detailed description of how each baseline was run.
  4. [Section III-D, Eq. (1)] Equation (1) does not appear to be a correct formulation of multiclass focal loss. With y_i denoting the predicted probability and p_i denoting the true distribution, the term (1-y_i)^gamma * p_i * log(y_i) would down-weight confident predictions and multiply by the target probability, which is not the standard focal loss that down-weights the contribution of easy normal-class examples. If this is a typesetting artifact, please provide the correct formula and define each variable precisely; if the loss is intentionally different, explain the design and its effect.
  5. [Section IV-B and IV-D] The manuscript omits standard training details needed to reproduce the results, including the optimizer, learning rate schedule, batch size, number of epochs, validation strategy, random seeds, and the number of model parameters. Since the paper makes a state-of-the-art claim on a public benchmark, these details are necessary for reproducibility and for assessing whether the comparison with baselines is controlled.
minor comments (6)
  1. [Author affiliation] The affiliation text contains a typo: 'Mateials' should be 'Materials'.
  2. [Section III-D heading] The heading 'D. Waveform-Logmel Audio Neural Networks' duplicates the title of Section III-C; this section appears to be about the loss function and should be renamed accordingly.
  3. [Figure 3 caption] The caption reads 'combined alone channel axis' and should read 'combined along the channel axis'.
  4. [Section IV-C] The metric equations (2)-(6) are badly garbled in the current typeset version (e.g., the numerator and denominator symbols appear corrupted). Please ensure that the formulas for sensitivity, specificity, average score, harmonic score, and total score are typeset correctly.
  5. [Throughout] The paper uses inconsistent notation for sensitivity: 'SE' in the abstract, 'Sen' in the tables, and 'SN' in the equations. Please standardize.
  6. [References] References [12] and [27] refer to the same paper by Chen et al. (BioCAS 2022) and should be merged or distinguished.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper is an empirical benchmark study with no fitted-as-predicted constants or self-referential derivation.

full rationale

The paper makes no theoretical derivation and presents no parameter fitted to data that is later renamed as a prediction. The core claim is an empirical performance comparison on the external SPRSound dataset, using standard components (1D-CNN, AST, Bi-GRU, focal loss). The only tuned hyperparameter mentioned is the focal loss focusing parameter gamma=2.0, which is a standard regularization choice and is not used to manufacture the reported sensitivity or total score; those numbers are measured on the test set. The comparison with prior work in Table II is an empirical benchmark, not a derivation. There is no self-citation that carries a load-bearing argument: the cited prior methods are external, and the architecture components are standard public techniques. Concerns about which SPRSound split produced Table II or whether micro-averaged binary metrics hide per-class failures are legitimate evaluation-protocol and correctness concerns, but they are not circularity: the result does not reduce by definition or by construction to its inputs. Therefore the appropriate circularity score is 0.

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

No new physical entities are introduced. The central claim rests on several hand-chosen hyperparameters and on the dataset and evaluation assumptions listed above. The focal loss gamma, the wavelet frame schedule, and the binary evaluation protocol are the most load-bearing choices because they directly shape the reported sensitivity values.

free parameters (5)
  • Focal loss focusing parameter gamma = 2.0
    Set in Section III-D to down-weight normal respiratory events. No sensitivity analysis or tuning procedure is reported, and the objective directly shapes the reported abnormal-class sensitivity.
  • 1D-CNN kernel and stride schedule = Kernel 80, stride 5 initially; kernel 80, stride 4 for three blocks
    Hand-designed in Section III-A to produce about 25 frames per second. This determines the temporal granularity of the waveform branch and has no ablation or justification.
  • Mel spectrogram configuration = 128 fbank bins, 25 ms Hamming window, 10 ms hop, 16 kHz resampling
    Standard audio feature choices from Section IV-B, but no ablation or dataset-specific rationale is given for the respiratory sound task.
  • AST patch size and overlap = 16x16 patches with overlap of 8 in time and frequency
    Patch geometry controls the token sequence and temporal resolution of the transformer branch. The paper states the choice in Section III-B without comparing alternatives.
  • Butterworth passband = 40 Hz to 850 Hz
    Chosen in Section IV-B to keep respiratory fundamentals and suppress heart sounds. This filtering affects both input branches and is not varied across experiments.
assumptions (3)
  • domain assumption The SPRSound annotations and official train/test splits are reliable ground truth for respiratory sound labels.
    All reported metrics inherit the dataset annotation quality and the split definitions from Section IV-A and Table I.
  • ad hoc to paper Micro-averaged binary sensitivity and specificity over pooled abnormal events is a valid proxy for multiclass respiratory sound classification performance.
    Equations (2) to (6) in Section IV-C collapse seven classes into normal versus abnormal, but the abstract and conclusion claim the model distinguishes pathological classes. No per-class evidence is provided.
  • ad hoc to paper Reshaping the 1D-CNN channel output into F groups by channel splitting produces a meaningful frequency dimension compatible with AST features.
    Section III-A states that splitting channels 'introduces frequency information' without a learnable or signal-processing basis. The fusion in Section III-C depends on this equivalence.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Waveform-Logmel Audio Neural Networks for Respiratory Sound Classification." pith.science (2026). https://pith.science/paper/PEBEY2KE

@misc{pith2026250417156,
  author       = {Pith},
  title        = {Pith review of: Waveform-Logmel Audio Neural Networks for Respiratory Sound Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PEBEY2KE}},
  note         = {Machine review of arXiv:2504.17156}
}
read the original abstract

Auscultatory analysis using an electronic stethoscope has attracted increasing attention in the clinical diagnosis of respiratory diseases. Recently, neural networks have been applied to assist in respiratory sound classification with achievements. However, it remains challenging due to the scarcity of abnormal respiratory sound. In this paper, we propose a novel architecture, namely Waveform-Logmel audio neural networks (WLANN), which uses both waveform and log-mel spectrogram as the input features and uses Bidirectional Gated Recurrent Units (Bi-GRU) to context model the fused features. Experimental results of our WLANN applied to SPRSound respiratory dataset show that the proposed framework can effectively distinguish pathological respiratory sound classes, outperforming the previous studies, with 90.3% in sensitivity and 93.6% in total score. Our study demonstrates the high effectiveness of the WLANN in the diagnosis of respiratory diseases.

Figures

Figures reproduced from arXiv: 2504.17156 by the authors.

Figure 1
Figure 1. The overall architecture of the Waveform [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 25 canonical work pages

  1. [30]

    An effective lung sound classification system for respiratory disease diagnosis using densenet cnn model with sound pre - processing engine

    Ma, Wei-Bang, et al. "An effective lung sound classification system for respiratory disease diagnosis using densenet cnn model with sound pre - processing engine." 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2022

  2. [1]

    Bender, Rose Grace, et al. "Global, regional, and national incidence and mortality burden of non -COVID-19 lower respiratory infections and aetiologies, 1990–2021: a systematic analysis from the Global Burden of Disease Study 2021." The Lancet Infectious Diseases (2024)

  3. [2]

    Ioana, et al

    Cristea, A. Ioana, et al. "Outpatient respiratory management of infants, children, and adolescents with post -prematurity respiratory disease: an official American Thoracic Society clinical practice guideline." American journal of respiratory and critical care medicine 204.12 (2021): e115-e133

  4. [3]

    Auscultation of the respiratory system

    Sarkar, Malay, et al. "Auscultation of the respiratory system." Annals of thoracic medicine 10.3 (2015): 158-168

  5. [4]

    This configuration effectively corresponds to capturing 20 ms of respiratory audio per frame and producing 25 frames of feature representations per second, enabling fine -grained analysis of respiratory sounds. B. Spectrogram-AST System The AST model has emerged as a pioneering architecture in the realm of spectrogram processing [15 ]. Structurally, the A...

  6. [5]

    Deep learning -based lung sound analysis for intelligent stethoscope

    Huang, Dong-Min, et al. "Deep learning -based lung sound analysis for intelligent stethoscope." Military Medical Research 10.1 (2023): 44

  7. [6]

    Artificial intelligence for heart sound classification: A review

    Chen, Junxin, et al. "Artificial intelligence for heart sound classification: A review." Expert Systems 41.4 (2024): e13535

  8. [7]

    Automatic Speech and Voice Disorder Detection using Deep Learning-A Systematic Literature Review

    Sindhu, Irum, and Mohd Shamrie Sainin. "Automatic Speech and Voice Disorder Detection using Deep Learning-A Systematic Literature Review." IEEE Access (2024)

Show all 30 references
  1. [8]

    Ast: Audio spectrogram transformer

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

  2. [9]

    Patch -mix contrastive learning with audio spectrogram transformer on respiratory sound classification

    Bae, Sangmin, et al. "Patch -mix contrastive learning with audio spectrogram transformer on respiratory sound classification." arXiv preprint arXiv:2305.14032 (2023)

  3. [10]

    Multi-View Spectrogram Transformer for Respiratory Sound Classification

    He, Wentao, et al. "Multi-View Spectrogram Transformer for Respiratory Sound Classification." ICASSP 2024 -2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024

  4. [11]

    Ast -sed: An effective sound event detection method based on audio spectrogram transformer

    Li, Kang, et al. "Ast -sed: An effective sound event detection method based on audio spectrogram transformer." ICASSP 2023 -2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023

  5. [12]

    Multiclass categorisation of respiratory sound signals using neural network

    Babu, Naseem, et al. "Multiclass categorisation of respiratory sound signals using neural network." 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2022

  6. [14]

    Convolutional recurrent neural networ ks for polyphonic sound event detection

    Cakır, Emre, et al. "Convolutional recurrent neural networ ks for polyphonic sound event detection." IEEE/ACM Transactions on Audio, Speech, and Language Processing 25.6 (2017): 1291-1303

  7. [15]

    Classification of lung sounds using CNN-Attention

    Li, Chenghan, Huaichang Du, and Bing Zhu. "Classification of lung sounds using CNN-Attention." EasyChair Preprint 4356 (2020)

  8. [16]

    Transformer -based approach to pathology diagnosis using audio spectrogram

    Tami, Mohammad, et al. "Transformer -based approach to pathology diagnosis using audio spectrogram." Information 15.5 (2024): 253

  9. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Dosovitskiy, Alexey, et al. "An image is worth 16x16 words: Transformers for image recognition at scale." arXi v preprint arXiv:2010.11929 (2020)

  10. [18]

    Heart sound classification based on improved MFCC features and convolutional recurrent neural networks

    Deng, Muqing, et al. "Heart sound classification based on improved MFCC features and convolutional recurrent neural networks." Neural Networks 130 (2020): 22-32

  11. [19]

    HCRNNIDS: Hybrid convolu tional recurrent neural network -based network intrusion detection system

    Khan, Muhammad Ashfaq. "HCRNNIDS: Hybrid convolu tional recurrent neural network -based network intrusion detection system." Processes 9.5 (2021): 834

  12. [20]

    Focal Loss for Dense Object Detection

    Lin, T. "Focal Loss for Dense Object Detection." arXiv preprint arXiv:1708.02002 (2017)

  13. [21]

    Sprsound: Open -source sjtu paediatric respiratory sound database

    Zhang, Qing, et al. "Sprsound: Open -source sjtu paediatric respiratory sound database." IEEE Transactions on Biomedical Circuits and Systems 16.5 (2022): 867-881

  14. [22]

    Generalized digital Butterworth filter design

    Selesnick, Ivan W., and C. Sidney Burrus. "Generalized digital Butterworth filter design." IEEE Transactions on signal processing 46.6 (1998): 1688-1694

  15. [23]

    Analysis of respiratory sounds: state of the art

    Reichert, Sandra, et al. "Analysis of respiratory sounds: state of the art." Clinical medicine. Circulatory, respiratory and pulmonary medicine 2 (2008): CCRPM-S530

  16. [24]

    Automatic adventitious respiratory sound analysis: A systematic review

    Pramono, Renard Xaviero Adhi, Stuart Bowyer, and Esther Rodriguez- Villegas. "Automatic adventitious respiratory sound analysis: A systematic review." PloS one 12.5 (2017): e0177926

  17. [25]

    Specaugment: A simple data augmentation method for automatic speech recognition

    Park, Daniel S., et al. "Specaugment: A simple data augmentation method for automatic speech recognition." arXiv preprint arXiv:1904.08779 (2019)

  18. [26]

    Understanding and using sensitivity, specificity and predictive values

    Parikh, Rajul, et al. "Understanding and using sensitivity, specificity and predictive values." Indian journal of ophthalmology 56.1 (2008): 45-50

  19. [27]

    Improving the resnet -based respiratory sound classification systems with focal loss

    Li, Jun, et al. "Improving the resnet -based respiratory sound classification systems with focal loss." 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2022

  20. [28]

    Classify respiratory abnormality in lung sounds using stft and a fine -tuned resnet18 network

    Chen, Zizhao, et al. "Classify respiratory abnormality in lung sounds using stft and a fine -tuned resnet18 network." 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2022

  21. [29]

    A feature polymerized based two -level ensemble model for respiratory sound classification

    Zhang, Lin, et al. "A feature polymerized based two -level ensemble model for respiratory sound classification." 2022 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 2022

  22. [31]

    TRespNET: A dual- route exploratory CNN model for pediatric adventitious respiratory sound identification

    TaghiBeyglou, Behrad, et al. "TRespNET: A dual- route exploratory CNN model for pediatric adventitious respiratory sound identification." Biomedical Signal Processing and Control 93 (2024): 106170. Method Architecture Sen (%) Spe (%) AS (%) HS (%) TS (%) Jun Li et al.[26] CNN ...

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.