Unsupervised speaker indexing using one-class Support Vector Machines

This paper addresses the unsupervised speaker change detection problem, which is a key issue in any audio indexing process. Here, we derive a new approach based on the Kernel Change Detection algorithm introduced recently by Desobry et al. This new algorithm does not require explicit modeling of the data, and is able to deal with large dimensional acoustic feature vectors. Several experiments using RT'03 NIST data show the efficiency of the algorithm. Comparisons to the well known GLR-BIC algorithm are presented, for various parameter settings.