The Bayesian information criterion (BIC) is a model selection criterion that has previously been applied to speaker segmentation of broadcast news by several researchers. The BIC approach treats speaker segmentation as a model selection problem. As the BIC requires the estimation of the sample covariance matrix, its performance tends to deteriorate as the speaker-turn duration decreases. It is well known that the BIC does not perform well on short segments, making the BIC inappropriate for conversational speech. In this paper, we estimate the hyperparameters of a prior distribution from a disjoint set of speakers and use the prior information to adapt the maximum a-posteriori distribution of the BIC. We show that this results in improved performance for a conversational telephone-speech corpus.
Cite as: Roch, M., Cheng, Y. (2004) Speaker segmentation using the MAP-adapted Bayesian information criterion. Proc. The Speaker and Language Recognition Workshop (Odyssey 2004), 349-354
@inproceedings{roch04_odyssey,
author={Marie Roch and Yanliang Cheng},
title={{Speaker segmentation using the MAP-adapted Bayesian information criterion}},
year=2004,
booktitle={Proc. The Speaker and Language Recognition Workshop (Odyssey 2004)},
pages={349--354}
}