Pass phrase based speaker recognition for authentication

Abstract

Speaker recognition in applications of our daily lives is not yet in widespread use. In order for biometric technology to make sense for real-world authentication applications and be accepted by end users, convenience of use, robustness and accuracy of such a system are equally important. This paper defines these requirements for pass phrase based voice authentication embedded within a multi modal biometric system and describes methods and algorithms developed and optimized for the demands of such an application. Classification is based on dynamic time warping which can cope with limited training data. MFCC features which have been optimized for speaker specific properties are used. Robustness of the system is increased with speech enhancement and cepstral mean subtraction. Furthermore, vector quantization with speaker specific codebooks is applied in order to decrease storage requirements for the biometric template. On an appropriate data base, a verification EER of 2.7% is achieved with limited training and test material.