关键词: 说话人检测, 改进贝叶斯信息准则, 检测精度, 可测度, 数据窗

Abstract:

In order to overcome such disadvantages as high computational cost and low detection precision existing in the speaker change detection with the Bayesian information criterion（BIC） algorithm,a modified BIC algorithm is proposed.In this algorithm,the maximum length of the first data window in the analysis window is restricted to reduce the computational cost,and the effective length,namely the detectability of the second data window in the analysis window is increased to improve the detection precision.Moreover, the growing computation without a long time speaker change is decreased by detecting potential speaker change points only in a new range. Experiment resuits show that, as compared with the conventional BIC algorithm, the proposed algorithm decreases the bias error from 0. 10 -0. 80 to 0. 03 - 0. 20, and saves 75 % of computation time for a 40-s analysis window.

Key words: speaker detection, modified Bayesian information criterion, detection precision, detectability, data window