An effective unsupervised scheme for multiple-speaker-change detection

This paper presents an enhanced Bayesian information criterion (BIC)-based algorithm for multiple-speaker-change detection (MSCD) without prior acoustic information on speakers. The enhancement offered by the proposed approach is in terms of effectiveness. This is achieved through the introduction of robustness into the standard BIC procedure, against certain important causes of misclassification. The paper also introduces a new measure, termed effective error rate (EFER), for evaluating the relative performance of MSCD algorithms. It is shown that the proposed measure allows a more meaningful evaluation of MSCD than the conventional ones. The experimental results obtained using this new evaluation measure clearly confirm the effectiveness of the proposed algorithm. The experimental investigation is based on 3 hours of broadcast news material with 445 speaker changes.