A new penalty term for the BIC with respect to speaker diarization

In this paper we revise the penalty term of the Bayesian Information Criterion (BIC). Based on our previous approach to penalize each cluster only with its corresponding effective sample size - which we called the Segmental-BIC - we examine a new formula of the penalty term. The criterion we derive has the appealing property of the Segmental-BIC, that is it approximates the evidence of overall partitions while leading to an autonomous pairwise dissimilarity measure. We tested our new criterion on two speaker diarization benchmarks and we report significant increase in accuracy.

• Type: Conference
• Year: 2010
• Where: ICASSP
• Authors: Themos Stafylakis, Georgios Tzimiropoulos, Vassilios Katsouros, George Carayannis
• Keywords: Bayesian Information Criterion, Corresponding Effective Sample, ICASSP 2010, Penalty Term, Signal Processing