Speeding Up Speaker Diarization by Using Prosodic Features

In this article we present a method to speed up agglomerative clustering used in speaker diarization by using long-term prosodic features. A set of these features is used to decide which clusters should be merged. This strategy reduces the number of decisions that have to be performed using the more calculation-intensive method based on the Bayesian Information Criterion (BIC). We show a speedup of 30 % to a state-of-the-art diarization system. This work was partly funded by DTO VACE contract number NBCHC060157. Gerald Friedland and Christian Müller were supported by a fellowship within the postdoc program of the German Academic Exchange Service (DAAD). Speeding Up Speaker Diarization by Using Prosodic Features Yan Huang,Gerald Friedland,Christian Müller, Nikki Mirghafori International Computer Science Institute, Berkeley Department of Computer Science, University of California, Berkeley {yan,fractor,cmueller,nikki}@icsi.berkeley.edu