Fusion of cross stream information in speaker verification

Abstract

This paper addresses the performance of various statistical data fusion techniques for combining the complementary score information in speaker verification. The complementary verification scores are based on the static and delta cepstral features. Both LPCC (Linear prediction-based cepstral coefficients) and MFCC (mel-frequency cepstral coefficients) are considered in the study. The experiments conducted using a GMM-based speaker verification system, provides valuable information on the relative effectiveness of different fusion methods applied at the score level. It is also demonstrated that a higher speaker discrimination capability can be achieved by applying the fusion at the score level rather than at the feature level.