Text-Dependent Speaker Verification System

Speaker verification technique is a very important branch of speech techniques. To verify a speaker's identity through his or her voice is a kind of biometric verification methods. This paper is mainly about techniques to complement a text-dependent speaker verification system. The system was divided into three parts: front end feature extraction part, speaker modeling and recognition part, and decision making part. Besides feature extraction, front end part also includes end point detection, channel normalization, feature selection. This study focused on feature selection algorithm and correlation beased feature selection algorithm is used. Compared with baseline system which had 26 features, when the number of selected features reached 18, the performance became better with the selection algorithm. When the number of selected features increased to 30, performance is much better and the feature set is very robust.Text-dependent speaker verification system can be implemented through different mechanisms. In the study, two kinds of systems were built: user-customized speaker verification system and text-promted speaker verification. In user-customized speaker verification system, we focused on building whole utterance model for every speaker and DTW,HMM and GMM modeling methods were adopted. The detail of DTW template matching was studied. In HMM based system, the number of HMM states was studied. Our text-prompted speaker verification system was based on Chinese digital set. And phoneme HMM was used as the basic acoustic models for speaker modeling.In decision making part, score normalization was studied in mathematical and physical aspects. Because training data was limited, our normalization method was in testing stage (like T-norm), and only top 20% scores was selected to compute the normalization score.