基于时变多示例学习的性别识别  

Gender identification based on time-varying multi-instance learning

作　　者：顾明亮[1,2] 张宁[1] 张世形 鲍薇[2] 

机构地区：[1]江苏师范大学物理与电子工程学院,江苏徐州221116 [2]江苏师范大学语言科学学院,江苏徐州221116

出　　处：《计算机工程与设计》2013年第11期4027-4031,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61040053);江苏省普通高校研究生科研创新计划基金项目(CXZZ11_0903;CXZZ12_0977;CX-LX12_0976)

摘　　要：为了提高性别识别(gender identification)的识别率,提出了一种基于时变多示例学习(multi-instance learning)的性别识别方法。该方法将语音段作为多示例包,语音的声学特征矢量经过K均值(K-means)聚类生成包中示例。将男、女性语音包标记成不同类别后,利用EM-DD(expectation maximization diverse density)算法求解出男、女性语音的多密度点,提出了Bags-K近邻分类算法进行识别。实验结果表明,性别识别系统平均识别率可达97%。A method for gender identification based on Time-varying Multi-Instance Learning is presented to improve the recognition rate. The method regards the speech regions as multi-instance bag, and the acoustic features of the speech regions are clustered into instances of the bag using K-means clustering algorithm. After labeling male and female speech bags into different kinds, EM-DD algorithm is used to get male and female speech diverse density points, and Bags-K neighbor classification algorithm is put forward for identification. Experiments show the average recognition rate of the gender identification system could reach as high as 97%.

关 键 词：多示例学习 性别识别 期望最大多样性密度 K均值 包-K近邻 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]