An Interpretable Four-Branch Wavelet Attention Network for Pathological Voice Detection
ID:192 View Protection:ATTENDEE Updated Time:2026-10-19 18:11:44 Hits:54 Poster Presentation

Start Time:Pending(Asia/Shanghai)

Duration:Pending

Session:No Session »

No files

Abstract
Abstract—Automatic detection of pathological voice can provide objective assistance for laryngeal disease screening, treatment follow-up, and telemedicine. Traditional methods typically rely on handcrafted acoustic features such as fundamental frequency, jitter, shimmer, harmonic-to-noise ratio, and mel-frequency cepstral coefficients, and their performance is susceptible to feature selection, recording conditions, and pathology variations. Deep learning can learn discriminative features directly from raw waveforms or time-frequency representations, but its internal decision process lacks clear interpretability. To address this issue, this paper proposes Four-Branch Wavelet Attention Network (FWANet). Parallel branches are constructed using db16, sym8, coif8, and rbio3.1 wavelets. Each branch alternates six stages of one-level discrete wavelet transform and one-dimensional convolution, while frequency-channel attention is applied in the last five stages to adaptively recalibrate low- and high-frequency features. After temporal alignment and sample-level branch weighting, a Cross-Wavelet Temporal Feature Fusion (CTFF) module integrates local temporal variation and cross-wavelet channel information. FWANet achieves accuracies of 0.9800, 0.7734, and 0.8545 on the MEEI, SVD, and HUPA databases, respectively, with MCC values of 0.9522, 0.5024, and 0.7153. Ablation results show that CTFF consistently improves Accuracy and MCC over the corresponding four-branch variant without CTFF. Branch weights and layer-wise frequency-channel attention further provide hierarchical interpretability.
 
Keywords
pathological voice classification; discrete wavelet transform; multilayer wavelet attention; frequency-channel attention; interpretable deep learning
Speaker
Le Jiang
student Soochow University

Submission Author
Le Jiang Soochow University
Fuming Ji Soochow University
Yiting Zhang Soochow University
MingXuan Yan Soochow University
xiaojun Zhang Soochow University
Zhi Tao Soochow University
Submit Comment
Verify Code Change Another
All Comments
Important Date
  • Conference Date

    Nov 06

    2026

    to

    Nov 08

    2026

  • Oct 15 2026

    Draft paper submission deadline

  • Nov 10 2026

    Registration deadline

Sponsored By
IEEE Instrumentation and Measurement Society
Organized By
Sichuan University