Voice pathology detection using machine learning algorithms based on different voice databases

The application of machine learning in analyzing voice disorders has become crucial for non-invasive voice pathology detection using voice signals. However, current systems face challenges such as low detection accuracy, limited databases, and evaluation metrics. More importantly, most existing studies rely on training and testing algorithms based on the same database, limiting their applicability in real-world scenarios with diverse data sources. Unlike traditional approaches that focus solely on single-database training and testing, this study presents a cross-database evaluation strategy to assess the robustness and generalizability of machine learning algorithms for voice pathology detection. Several algorithms, including Online Sequential Extreme Learning Machine (OSELM), Support Vector Machine (SVM), Decision Tree (DT), and Na & iuml;ve Bayes (NB), were evaluated using two databases: the Saarbrucken Voice Database (SVD) and the Malaysian Voice Pathology Database (MVPD). Two scenarios were considered: (1) training and testing on the same database and (2) training on one database and testing on another. The proposed study uses the Mel-Frequency Cepstral Coefficient (MFCC) technique for extracting features from voices. The algorithms are assessed using many evaluation metrics such as accuracy, precision, sensitivity, specificity, F-measure, and G-mean. Experimental results demonstrate that the OSELM algorithm achieves superior performance across both scenarios, with accuracies of up to 85.71 % in Scenario 1 and 80.77 % in Scenario 2, outperforming other algorithms. This novel approach highlights the reliability of OSELM and the importance of cross-database testing for developing robust and generalizable voice pathology detection systems.