Dynamic speaker clustering algorithm based on minimal GMM distance tracing

Abstract

In the field of speaker clustering, most of the clustering algorithms rely heavily on the pre-given thresholds, which is hard work to get the optimal values. This paper proposed a speaker clustering algorithm based on tracing the minimal Bhattacharyya distance between two Gaussian Mixture Models (GMMs), without any pre-given thresholds. In the procedure of clustering, if utterance set A and B has the minimal distance, utterance B is regarded as suspicious set whose utterance may come from the speaker of A. And then, two stage-verification is used. First, a comparative likelihood is used to verify whether the suspicious set B is generated from the speaker or not. Second, a comparative likelihood for each utterance in set B is used to judge whether it is produced by the speaker of set A or not. If the utterance is from the speaker of set A, we move the utterance of set B to set A. And then the models of utterance set A and B are updated. Repeat the above two stages until each speech set is not changed. Experiments, evaluated on Chinese 863 speech database, give 68.97% average cluster purity (ACP), and classification error ratio (CER) is 39%. On the other hand, CER of the K-means and the Iterative Self-Organizing Data Analysis (ISODATA) with the optimal thresholds give 35% and 38% respectively.