A text-independent speaker recognition system based on Probabilistic Principle Component Analysis

To alleviate the problem of severe degradation of speaker recognition performance because of the phoneme variability between training and testing speech data, in the text-independent speaker recognition system. The paper proposed a text-independent (TI) speaker identification method that suppresses the phonetic information by a subspace method, Probabilistic Principle Component Analysis (PPCA) is utilized to construct these subspaces. Firstly, the covariance matrix was obtained from the large training speech feature data, and then the projection matrix was obtained using the EM algorithm. In the proposed method, it is assumed that a subspace with large variance in the speech feature space is a “phoneme-dependent subspace” and a complementary subspace of it is a “phoneme-independent subspace”, the feature vectors of train/test speech data are projected to a phoneme-independent subspace and a new feature vectors are obtained. In GMM-based TI speaker identification experiments, the new feature vectors improves the identification rate by 16.25% and 2.99% respectively, compared with conventional MFCC, PCA-based MFCC. It shows that the new feature vectors of the proposed method can efficiently capture speaker-discriminative information, and suppress the other speech information.