Hierarchical speaker identification using speaker clustering

We explore 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 to a distance measure. The time spent on HSI (hierarchical speaker identification) is approximately 30.3 percent more than that spent on CSI (conventional speaker identification) when the number of registered speakers is 40 in our experiments. Increasing of the number of speakers decreases the time spent on HSI compared with CSI. It is shown that this approach can improve the speed of the speaker identification system.