Speaker recognition system using the improved GMM-based clustering algorithm

According to the sensitivity of initial value in the traditional GMM-based clustering algorithm, this paper proposes the improved GMM-based clustering algorithm which utilizes subtractive clustering to initialize the cluster centroids and uses an approximating K-L divergence as the distance measure. After clustering the universal background model(UBM) is trained for each clustering. In speaker recognition, the algorithm firstly confirms which clustering the aim speaker belongs to and then it uses the value of maximum likelihood probability and the UBM-based testing approach to recognize. According to the results of simulation using Matlab, the improved GMM-based clustering algorithm has the higher clustering accuracy and recognition rate than the traditional GMM-based clustering algorithm.