基于矢量量化的时序说话人聚类方法  

Sequential Speaker Clustering Based on Vector Quantization

作　　者：陈祝允 李艳雄[1] 杜佳媛 

机构地区：[1]华南理工大学电子与信息学院,广州510640

出　　处：《科学技术与工程》2014年第2期41-44,共4页Science Technology and Engineering

基　　金：国家自然科学基金项目(61101160);广州市珠江科技新星专项(2013J2200070);中央高校基本科研业务费专项资金重点项目(2013zz0053);国家级大学生创新训练项目(201210561046);广东省大学生创新训练项目(1056112028)资助

摘　　要：针对传统分层聚类方法运算速度较慢的问题,提出一种基于矢量量化的时序说话人聚类方法。首先对各语音段的特征进行矢量量化得到各语音段的码本,然后采用贝叶斯信息判据计算各码本之间的距离,最后按时间先后顺序进行说话人聚类。采用会议和新闻语音数据进行测试,实验结果表明:会议语音的说话人聚类F值为73.47%,新闻语音的说话人聚类F值为80.00%;在处理速度方面,该方法比无矢量量化时序聚类方法提高了3.16倍,比传统分层聚类方法提高了53.31倍。In view of slow speed of the traditional hierarchical clustering method, a method of sequential speak- er clustering based on vector quantization is proposed. First, vector quantization is used for obtaining codebooks by coding features extracted from speech segments. Then, according to Bayesian information criterion, it calculates the BIC distances between any two eodebooks. Finally, sequential speaker clustering is carried out. The experimental results, by testing meeting and news speech data, show that： the proposed method obtains F-score of 73.47% for meeting speech data and 80. 00% for news speech data. In terms of processing speed, the proposed method speeds up 3.16 times compared to the sequential clustering method without vector quantization, and 53.31 times compared to the traditional hierarchical clustering method.

关 键 词：时序说话人聚类 矢量量化 贝叶斯信息判据 梅尔频率倒谱系数 

分 类 号：TN912.3[电子电信—通信与信息系统]