Processing and Classification of Covid-19 Audio Data with Optimization-Based Learning Techniques
Yüsra Dağılma, Erdal Özbay · 2024
This study compares and evaluates the performance criteria of the techniques used to identify Covid-19 illness using cough sound signals. The study recommends highperformance methods for future research. A dataset containing cough sounds was used in the study. 12 different feature extraction methods were used to significantly improve the performance. MFCC, Chroma, Mel-spectrogram, Spectral contrast, Tonnetz, SC, Spectral roll off, ZCR, Spectral bandwidth, RMSE, Spectral flatness, and Pitch feature extraction techniques were used to extract important features from sound signals. All feature extractions were combined to obtain a total of 1224 features. After feature extraction, SMOTE algorithm was applied because there was data imbalance. The number of positive classes with 125 data was increased to 960. Negative and positive class values were set to 960. Data imbalance was eliminated with the SMOTE application. Then, Adam optimization was applied together with two different classification methods. The classification was performed by applying Adam optimization to CNN and MLP modeling. As a result of the classification, MLP gave an accuracy value of approximately $92.65 \%$, while CNN gave an accuracy value of approximately $84.93 \%$. It was observed that MLP showed higher and more successful performance than CNN.