Score Fusion in Text-Dependent Speaker Recognition Systems

摘要

According to some significant advantages, the text-depen dent speaker recognition is still widely used in biometric systems. These systems are, in comparison with the text-independent, more accurate and resistant against the replay attacks. There are many approaches regarding the text-dependent recognition. This paper introduces a com bination of classifiers based on fractional distances, biometric dispersion matcher and dynamic time warping. The first two mentioned classifiers are based on a voice imprint. They have low memory requirements while the recognition procedure is fast. This is advantageous especially in low cost biometric systems supplied by batteries. It is shown that using the trained score fusion, it is possible to reach successful detection rate equal to 98.98 % and 92.19 % in case of microphone mismatch. During verifica tion, system reached equal error rate 2.55% and 6.77% when assuming the microphone mismatch. System was tested using Catalan database which consists of 48 speakers (three 3 s training samples per speaker).