Kernel Function and Dimensionality Reduction Effects on Speaker Verification System

In this research work, we are dealing with an Automatic Speaker Verification problem, which consists on the determination if a person really is the person he/she claims to be, using his/her speech signal characteristics. Therefore, we conducted three series of experiments applied to a subset of Hub4 Broadcast-News database using a Support Vector Machine (SVM) as classifier and the Mel Frequency Spectral Coefficients (MFSC) as speakers’ features. In the first series of experiments, we investigated the effect of the number of features in order to obtain the minimum number that gives a Good Verification Score (GVScore). In the second series of experiments, as there are many types of kernel functions, the appropriate one for speaker verification is investigated. In the last series of experiments, we investigated the GVScore with regard to speaker gender (Male vs. Male, Female vs. Female and Male vs. Female). In our approach, we have used the MFSC as features extraction, which are calculated in both training and testing sessions. The obtained results of the proposed techniques are quite interesting.