Voice Pathology Detection Using Support Vector Machine Based on Different Number of Voice Signals
Fahad Taha AL‐Dhief, N. M. Abdul Latiff, Marina Mat Baki, Nik Noordini Nik Abd Malik, Naseer Sabri, Musatafa Abbas Abbood Albadr
Abstract
In voice pathology detection system, machine learning algorithms play an important role in the classification process. Most of the developed systems in voice pathology detection used a fixed number of voice signals. However, the performance of such systems may be affected by the increasing or decreasing number of voice signals. In other words, the system may not achieve the desirable detection accuracy if the number of voice signals used in the system changes. This paper presents a voice pathology detection system based on a different number of voice signals. In this work, the voice signals for the vowel /a/ of healthy and pathological classes are taken from the Saarbrücken voice database (SVD). The features of voice signals are then extracted using Mel-Frequency Cepstral Coefficient (MFCC). For the classification part, the Support Vector Machine (SVM) is used to classify the voice signals into healthy or pathological. Furthermore, the proposed system is evaluated in terms of accuracy, specificity, and sensitivity. The experiments results show that the highest achieved accuracy, specificity, and sensitivity for the proposed SVM are 84.37%, 90.90%, and 80.95%, respectively.