Using confidence measures to evaluate the speaker turns in speaker segmentation

In this paper, we propose a speaker segmentation algorithm using confidence measures, named CM-DISTBIC, which inserts a confidence score computation and fusion procedure into the two-step DISTBIC and MDISTBIC. In the first step, symmetric Kullback-Leibler distance (KL2) distance is replaced by Bayesian information criterion (BIC) distance to obtain a lower misdetection rate. In the second step, three different confidence measures are attached to the speaker change candidates according to the distance curve derived from the first step. False alarm peaks with relatively low fused confidence scores are eliminate from the set of potential speak turns. In the third step, speaker turn candidates are validated through BIC criterion. Compared with DISTBIC and MDISTBIC, the CM-DISTBIC conducted on the broadcast news corpora receives an increase of more than 11.5% and 8.9% in F-score respectively.