Hierarchical speaker identification using speaker clustering

This paper explores an approach to speaker identification called speaker clustering in the GMM-based speaker recognition system in order to reduce the computational complexity. The ISODATA algorithm adapted for our purpose works well when we cluster speakers whose acoustic characteristics are similar with a distance measure. The time spent on HSI (hierarchical speaker identification) is approximately 30.3 percent than that spent on CSI (conventional speaker identification) when the number of registered speaker is 40 in our experiments. And with the increasing of the number of speakers time spent on HSI decrease comparing with CSI. It is shown that this approach can improve the speed of speaker identification system for practical purpose.