Generalized cosine similarity in I-vector based automatic speaker recognition systems

This paper deals with the problem of processing of I-vectors in the text-independent speaker verification systems. A new generalized cosine similarity optimization technique is proposed. The optimization is performed over sets of orthogonal and diagonal matrices. Lie group techniques are used in order to obtain exact orthogonal matrix solution. The experiments have been performed using the NIST 2010 database. It was shown that a substantial equal error rate (EER) reduction can be obtained with the proposed algorithm.