基于特征均值距离的短语音段说话人聚类算法

Feature Mean Distance Based Speaker Clustering for Short Speech Segments

该文提出一种基于特征均值距离的短语音段说话人聚类算法。首先,定义特征均值距离用来在特征层而不是模型层刻画两个类之间的相似度;然后,迭代合并特征均值距离最小的两个类,直到任意两类之间的特征均值距离的最小值大于一个自适应门限为止。采用取自两个语音数据库的短于3 s的语音段进行实验测试,结果表明:与基于AHC+BIC的算法相比,F度量值平均提高了5%,运算速度约为以前算法的4.68倍。

An algorithm of speaker clustering is proposed based on Feature Mean Distance（FMD） for short speech segments.First,a distance measure,i.e.FMD,is introduced to represent the similarities between two clusters on the level of feature instead of the level of model.Then,two clusters with the minimum of FMDs are iteratively merged until the minimum of FMDs is larger than an adaptive threshold.Experimental results show average 5% improvements in F measure are obtained in comparison with the AHC＋BIC based algorithm.In addition,the proposed algorithm is 4.68 times faster than the AHC＋BIC based algorithm.