Genetic learning of multi-attribute interactions in speaker verification

Genetic algorithms are applied to identify the interactions of multiple speech features, represented by fuzzy measures, for speaker recognition. This work aims to investigate more thoroughly the use of fuzzy measures and fuzzy integral in information fusion by means of genetic optimization. The proposed approach is implemented into the speaker verification system and tested against a commercial speech corpus. The results in terms of equal error rates show that the proposed speaker verification system is more favorable than the conventional normalization, and lambda-measure fuzzy-integral based methods.