Speaker Recognition via Similar Centroids Balanced FKVQ

To solve the false acceptance problem in speaker recognition, definition of similar centroids among all reference speakers' codebooks and similarity between each two speakers are given. A triple is introduced to mark each centroid with the description of similar centroids. And such triple is used to modify the target function in the fuzzy kernel vector quantization speaker recognition, in which the distortion between a test speech feature vector and a centroid is balanced by the centroid's all similar centroids. The proposed algorithm can weaken similar centroids' effect in recognition, and decrease false acceptance rate. Similar centroids balanced algorithm showed excellent effect in speaker recognition, and especially in the case of reference speakers with high similarity.