Robust unsupervised speaker turn detection

In this paper we address an aspect of speaker recognition task, viz. unsupervised speaker turn detection. A metric based approach with two-pass criteria is proposed for this task. A GMM-based modified Log Likelihood Ratio metric is used in the first pass; Bayesian Information Criterion (BIC) metric is used in the second pass to verify or discard the speaker turn points hypothesized in the first pass. We consider two cases: long speaker turn segments (> 2 sec.) and short speaker turn segments (< 2 sec.). We have evaluated our algorithm using TIMIT speech files. Our precision results range from 85% to 93%, recall ranges from 75% to 78%, and the F-ratio is in the range 80-85%. These results are better than what has been reported in the literature so far.