Text Dependent Speaker Authentication System

Speaker authentication is a kind of technology which can verify users' entities by computers automatically. It can be devided as two parts: voiceprint verification and content verification. This dissertation mainly focuses on channel robust feature extraction, speaker verification and verbal information verification over telephone channel. Including:1. In speaker verification over telephone, the author firstly studies some common channel normalization algorithms. It's shown in the experiment that after using RASTA filter followed by cepstral variance normalization, the PLP feature can reduced from channel noise. This kind of feature will be used as our baseline system. Then feature selection algorithm based on corcorrelation information is proposed to represent the best feature set. Finally the selected features will be transformed to the discriminative feature space using linear discriminate analysis. It is shown by the experimental results that after these two steps false rejection rate of the system will be reduced by 17.4% compared with the baseline system.2. In this paper, three traditional algorithms: VQ, DTW and HMM are used in the text dependent speaker verification. Besides, the background VQ codebook is proposed as models for DTW and VQ methods. The experimental results show that the false rejection rate is reduced by 18.6% and 27.0% respectively using codebook normalization method. This paper also integrates these three system and the systems based on pitch, energy, duration together by mulitiple layer neural networks followed by AdaBoost machine learning algorithms. The performance will be futher improved with enhanced classifiers.3. Verbal information verification technology verifies the speech signal with personal transcriptions in order to judge whether the unknown speaker has the knowledge of the claim user. Utterance verification is a general solution for this technology. In this paper, confidence measures based on multiple levels of acoustic layers and models are proposed to give the verification scores. Finally the system fuses these scores to neural networks followed by AdaBoost to enhance the classifer. Compared with a single classifier, the relative equal error rate reduction is 23%. In addition this paper also combines zero normalization and context threshold normalization methods together, the performance is futher improve from 1.39% to 0.58% in terms of equal error rate.4 . This paper studies the application solutions of speaker authentication system, also discusses the relationship of password contents, voiceprint and security. Finally speaker authentication system combines both speaker verification and verbal information verification technologies. The experimental results show that when the imposters don't know the password of claim speakers in just one question set the system can achieve a false alarm of 0.02% and correct accpectance of 98%.