This paper proposes normalisation methods based on fuzzy set theory for speaker verification. A claimed speaker's score used to accept or reject this speaker is viewed as a fuzzy membership function. We propose two scores: the fuzzy entropy and fuzzy C-means membership functions. Moreover, a likelihood transformaiton is considered to obtain a general approach and, based on this, five more fuzzy scores are proposed. Finally, a noise clustering method is applied to the current and proposed methods, reducing the equal error rate in all cases. Experimetns performed on the ANDOSL and YOHO speech corpora show better results for all proposed methods.
展开▼
