Trajectory Model And The Vq Model-based Speaker Recognition Study

This thesis studies Trajectory Model and Vector Quantization Model in speaker recognition. Speaker recognition is a biometrics that recognizes people via their voice, and people focus it because it's a convenient, economical and accurate method. Trajectory Model was proposed recently in speech recognition, and it can explicitly model the dynamic information in the successive speech frames. VQ model describes speaker characteristics with the speech feature centroids and it can't represent the correlation in the speech feature frames. Now we introduce trajectory model in speaker recognition and improve the codebook training algorithm for VQ.The author discusses a speaker recognition system in many aspects, including feature extraction, modeling, patter match and decision. Some improvement have made based on the research.1. Segment Model: Temporal correlation being in the successive speech frames often contains the useful information that can characterize the speaker. The traditional feature only considers the correlation of intraframe and ignores the correlation of interframe. The speech signal is essentially continuous, so the speaker recognition rate can be improved if we obtain the temporal correlation in the frames. Segment Model can help us get the correlation.2. Parametric Stochastic Trajectory Model: In a speaker recognition system, it's often encountered that the speech data isn't enough for training. In this case, text-independent speaker recognition will result in a bad speaker recognition rate with GMM or HMM. An alternative is text-dependent speaker recognition. I implement the parametric stochastic trajectory model in text-dependent speaker recognition; and because of the strong ability in exploiting the time correlation of speech frames, the result is superior to those attained by GMM and CHMM.3. VQ Model Based on Clustering Validity Analysis: it's the flaw for the codebook training algorithm that the size of codebook is choosed artificially. Now we uses cluster validity analysis for refrence and improve it. The new codebook training algorithm can choose the size of codebook automatically and improve the performance of speaker recognition.