Sequential speaker clustering based on second order statistical measures

This paper presents an investigation on the sequential clustering, based on second order statistical measures, for the task of speaker diarization. This clustering technique, which represents the second part of a global system of speaker indexing, is made for gathering the homogeneous segments obtained during the segmentation step. Most existing clustering systems are based on hierarchical clustering as agglomerative techniques. Such systems present two problems: finding the stopping criterion and choosing the threshold of clustering decision. In this research work, we try to resolve the first problem (stopping criterion) by proposing a sequential clustering approach based on second order statistical measures in order to gather the similar homogeneous segments (obtained by the segmentation process). At the end, each class will contain the global intervention of only one speaker in the entire audio document. Our sequential clustering approach uses a mono gaussian measure, called μ
_Gβ
, which is able to assess the degree of similarity between the different homogeneous segments. Experiments are done on a subset of Hub4 Broadcast-News database and the corresponding results show that the implemented algorithms are interesting for the task of speaker clustering.