摘要
说话人聚类是说话人分离中的一个重要过程,然而传统的以贝叶斯信息准则作为距离测度的层次聚类方式,会出现聚类误差向上传递的情况。本文提出了一种逐级算法增强处理机制。当片段之间的最小贝叶斯信息准则距离超过设定的门限值时,或者类别个数到达一定程度时,将当前聚类结果作为初始类中心,通过变分贝叶斯迭代法重新对每个类别中的片段调优,最后再依据概率线性判别分析得分门限确定说话人个数。实验表明,本文方法在美国国家标准技术署08summed测试集上,使得"类纯度"和"说话人纯度"比传统算法都有了一定提升,且使得说话人分离整体性能相对提升了27.6%。
The speaker clustering is an important process of speaker diarization,yet traditional method for hierarchical agglomerative clustering(HAC)with distance measurement based on Bayesian information criterion(BIC)can lead to the clustering error propagation.To solve this problem,step by step algorithm is proposed,when the minimum BIC distance between segments exceeds a predefined threshold,or the number of the categories on hierarchical clustering reaches a certain number.The current clustering result as the initial class center,and then variational Bayesian method will be exploited to tune the speaker segments among the categories iteratively.Finally,the number of speaker is determined according to the probabilistic linear discriminant analysis(PLDA)score threshold.Experiments on national institute of standards and technology(NIST)08summed test set show that this method improves the "class purity" and "speaker purity" compared with conventional algorithms.Moreover,performance of speaker diarization is relatively improved by 27.6%.
出处
《数据采集与处理》
CSCD
北大核心
2017年第1期54-61,共8页
Journal of Data Acquisition and Processing
基金
公安部应用创新计划(2014YYCXGAES048)资助项目
关键词
说话人聚类
贝叶斯信息准则
概率线性判别分析
变分贝叶斯
speaker clustering
Bayesian information criterion
probabilistic linear discriminant analysis
variational Bayesian