Robust lightweight COVID-19 cough audio classification under limited data conditions

Abstract

The rapid global spread of COVID-19 has highlighted the need for efficient and scalable respiratory disease screening methods. Compared with conventional diagnostic approaches such as RT-PCR, cough sound–based analysis provides a non-invasive and low-cost alternative. This paper proposes a lightweight deep learning framework for cough-based respiratory disease classification using log-Mel spectrogram representations. A five-layer convolutional neural network is employed as the backbone to enable efficient learning under limited-data conditions. The influence of spectral resolution is investigated by evaluating different Mel-spectrogram configurations, while data augmentation and test-time augmentation (TTA) are incorporated to improve robustness and inference stability. Experimental results demonstrate that the proposed framework achieves consistent and competitive performance across multiple evaluation metrics. In particular, robustness enhancement strategies significantly improve generalization performance, while compact spectral representations remain effective for discriminative feature learning. The results indicate that combining lightweight architectures with appropriate feature representations and robustness-oriented strategies provides an effective and practical solution for cough-based respiratory disease screening.