摘要： 针对传统T2-BIC算法累积误差较大、召回率不高的问题，提出一种改进的T2-BIC说话人二级分割算法。第1级采用改进的滑动窗口检测搜索窗中的T2统计量峰值，利用贝叶斯信息准则(BIC)对峰值进行确认，第2级利用分步解决的思想处理由于BIC可信度过低而漏选的分割点。实验结果表明，与同类算法相比，该算法分割效果较好，准确率、召回率和综合性能都有所提高。

关键词: T2统计量, 贝叶斯信息准则, T2-BIC算法, 分步解决

Abstract: This paper proposes an improved two-stage T2-BIC algorithm for speaker segmentation, because traditional T2-BIC algorithm has the problems of a bigger accumulated error and a lower recall ratio. In the first stage, the peak position of T2 statistic in search window is detected by using improved sliding variable-size analysis window, and Bayesian Information Criterion(BIC) algorithm is used to acknowledge the peaks. In the second stage, the idea of divide-and-conquer is used to detect the missed turns because of low BIC reliability. Experimental result shows that compared with other algorithms, the improved algorithm achieves better performance, and improves the precision, recall and F measure.

Key words: T2 statistic, Bayesian Information Criterion(BIC), T2-BIC algorithm, divide-and-conquer

中图分类号: 

• TP391