Verbal Information Verification for High - performance Speaker Authentication

of thesis entitled: Verbal Information Verification for High-performance Speaker Authentication Submitted by QIN Chao for the degree of Master of Philosophy in Electronic Engineering at The Chinese University of Hong Kong in June 2005. Automatic speaker authentication is to authenticate the identity of a claimed speaker by verifying the identity-related information embedded in his/her spoken utterances. The information mainly refers to the voice characteristic and the verbal content. This leads to two closely related technologies. One is speaker verification (SV) and the other is verbal information verification (VIV). Throughout this thesis, speaker authentication is studied for users who speak Cantonese. A SV system has been developed using state-of-the-art techniques with digit-based contexts. The universal background model (UBM) adaptation method are adopted and compared with the Cohort method. In VIV, many considerations in the process of constructing anti-models, e.g., the way of pooling subword units, the model structure and the model complexity, are discussed. A new technique is proposed to provide more reliable and effective anti-likelihood scores. Our method uses the Gaussian Mixture Model (GMM) instead of the conventional Hidden Markov Model (HMM) for anti-modeling at the subword level. Three methods for integrating SV and VIV systems are investigated. They are voting method, support vector machines (SVMs) and Gaussian-based classifier respectively. Experiments on 20 Cantonese speakers show that for SV the best result with equal error rate (EER) of 1.5% can be attained using the UBM adaptation. For VIV the experimental results indicate that the proposed GMM-based anti-