基于改进Citation-KNN算法的性别识别研究  

Study of speaker gender identification based on modified Citation-KNN

作　　者：朱俊梅[1] 顾明亮[1,2] 张世形 贾晶晶[1] 

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

出　　处：《计算机工程与应用》2015年第15期206-210,共5页Computer Engineering and Applications

基　　金：国家自然科学基金(No.61040053);江苏省普通高校研究生科研创新计划项目(No.CXZZ11_0903;No.CXLX12_0976;No.CXZZ12_0977)

摘　　要：为了简化系统模型训练方法,提高性别识别系统的整体效率,提出了一种基于改进Citation-KNN算法的说话人性别识别方法。该方法将连续语音切分,训练每段语音的高斯混合模型(Gaussian Mixture Model,GMM)作为多示例包,其所有混合元为相应包中示例;采用改进的Hausdorff距离作为包与包之间的距离测度,通过Citation-KNN算法进行性别识别。该方法以多示例包间距离为分类依据,简化了系统训练,且识别率优于一些传统算法。To simplify the method of model training and improve efficiency of gender identification system, an algorithm of speaker gender identification based on modified Citation-KNN algorithm and a new speech Multi-Instance（MI）bag generating method are introduced. Continuous speech is segmented and sub-paragraph speech is modeled with a Gaussian Mixed Model（GMM）. The generated GMM is treated as an MI bag and the components of the GMM are the instances of the corresponding bag. Modified Hausdorff distance is used to measure the distance between two bags. Experimental results show that this method can identify speakers＇ gender effectively and is superior to traditional algorithms in the correct recognition rate.

关 键 词：性别识别 改进Citation-K最近邻(KNN)算法 高斯混合模型 改进HAUSDORFF距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]