A combination of Gaussian Mixture Model and Support Vector Machine for speaker verification

In this paper, we proposed a speaker verification system to determine whether an input speech comes from outside the set of known speaker robustly. The proposed system consists of preprocessing, feature extraction, distortion measure calculation, and verification stages. The proposed speaker verification firstly catches and segments speech in the preprocessing stage. The segmented speech is extracted to MFCC feature, known as the most popular feature in speech processing, and a Gaussian Mixture Model (GMM) is constructed to model the extracted feature vectors. Next, a high dimensional distance between it and GMM, which is model of pre-trained speech of claimed identity, is calculated as a multi-scoring vector. Finally, a support vector machine decides whether the distance is acceptable or not, by other words, the input speech is verified or rejected. Experiment results show that the proposed system can recognize the claimed speaker with an accuracy of 96%, while the error rate is 6.6% acceptable.