Speaker identification based on state space model

Abstract

The dominant model for speaker recognition keeps Gaussian Mixture Model (GMM), though hidden Markov model (HMM) is successfully used in speech recognition. And then, few attentions have been paid on the correlation of features when modeling in the speaker recognition system, which may improve the final performance. Coincidentally, state-space model (SSM) is a model relying on the features correlation. A novel speaker identification method using a SSM based on the speaker features is proposed. The SSM, which has never been used in speaker identification before, presents the advantage of explicitly modeling spectral parameter trajectory dynamically. The preliminary system in our paper, directly regarding Mel frequency cepstrum coefficient (MFCC) as features, was set up to prove the correctness of SSM using in speaker recognition. The senior experiments, using supervectors, further explained the importance of the inner relationship of the speaker features, since considering SSM into GMM-UBM supervectors shows better accuracy, although just winning about 1 % in Arabic digital experiment.