[Legacy Report] Generalized cosine similarity in I-vector based speaker recognition systems
A problem of processing of I-vectors in the text-independent speaker verification systems was discussed. A new generalized cosine similarity optimization technique was proposed. The optimization was performed over sets of orthogonal and diagonal matrices. Lie group techniques were used in order to obtain exact orthogonal matrix solution. The experiments have been performed using the NIST2010 database. It was shown that a substantial equal error rate (EER) reduction can be obtained with the proposed algorithm.