A Deep Learning Model for Detecting COVID-19 Cough Using MFCC And Mel Feature Fusion

Accurate detection of respiratory diseases plays a critical role in pandemic containment strategies. In this paper, we propose a dual-branch deep neural architecture specifically optimized for COVID-19 detection in terms of binary classification through acoustic analysis. The framework employs an ECAPA-TDNN encoder to process two complementary spectral representations, namely, Mel-frequency cepstral coefficients (MFCC) spectrograms and Mel spectrograms. Subsequently, the architecture incorporates coordinated local-global feature extraction modules that perform hierarchical fusion of features. To address class imbalance challenges inherent in respiratory acoustic datasets, we introduce a weighted composite loss function. Experimental validation on the Coswara dataset demonstrates superior detection efficacy of the proposed model, outperforming existing benchmark methods.