Comparison of speaker segmentation methods based on the Bayesian Information Criterion and adapted Gaussian mixture models

This paper addresses the topic of unsupervised speaker segmentation for automatic speech recognition in a complex real life environment like broadcast news domain. A statistical approach where a Universal Background Model (UBM) is applied for online speaker segmentation was compared with the widely used Bayesian Information Criterion (BIC) approach. An analysis of influence of different window selection strategies on performance of both methods was carried out. Experiments and test evaluation were performed on the Slovenian BNSI Broadcast News speech database.