Grid-density based feature classification for speaker recognition

Abstract

A new strategy of feature classification method for speaker recognition based on the grid-density clustering is presented. According to the concept of density-based and grid-distance-based distribution in the Mel-frequency cepstrum domain, the feature vectors of each speaker were self-adaptively classified into L clusters with less overlapped. With these convex and non-interwoven clusters, the Gaussian Mixture Model could statistically analyze and estimate the distinct feature classification for each speaker. Moreover, a new speaker recognition system was established by using GMM-UBM model. The experimental results showed that the clustering effect of the proposed method was superior to the K-means plus EM clustering method, and the proposed speaker recognition system achieves better classification performance in terms of verification accuracy and computational complexity.