A Speaker Verification System for Arabic

Abstract

In this thesis a new Speaker Verification System (SVS) methodology is developed. The objective is to develop a SVS system that achieves high identification accuracy with low storage requirements for the use in smart card-based systems. This proposed system combines the advantages of the simple structure of text-dependent systems and the performance of text-independent systems. The feature extraction utilized the Mel-Frequency Cepstrum Coefficients MFCC for its superior performance, and user template is based on a Vector Quantization codebook for fast and efficient pattern classification. A new adaptive threshold is also proposed to achieve improved performance with time and to mitigate temporarily changes in speaker voice. The system is trained and tuned for Arabic speakers using initially certain parts front the holy Qura'an. A sample of eighteen Arabic speakers is used to evaluate the system performance. The system achieves an accuracy of 0.01 for the False Acceptance Rate (FAR), and 0.04 for the False Rejection Rate (FRR).