基于共性特征选择的短时说话人识别方法  

Short Time Speaker Recognition Method Based on Common Feature Selection

作　　者：肖星星[1] 冯瑞[1] 

机构地区：[1]复旦大学计算机科学技术学院,上海201203

出　　处：《计算机工程》2012年第24期171-174,178,共5页Computer Engineering

基　　金：上海市教育委员会科研创新基金资助项目(11CXY01;11CXY03)

摘　　要：现有说话人识别方法在短时语音条件下识别性能明显下降。为此,提出一种基于共性特征选择的短时说话人识别方法。利用说话人语音数据得到高斯混合模型,提取说话人之间的公共重叠部分,建立共性重叠模型和非重叠模型,根据这2个模型完成测试语音特征的选择,计算其在所有说话人非重叠模型中的相似度,并根据相似性最大化原则进行决策。实验结果表明,该方法具有较强的鲁棒性,且系统识别错误率较低。Under the condition of short time,the performances of the existing methods of speaker recognition decrease apparently.To solve this issue,this paper proposes an algorithm of short time speaker recognition method based on common feature selection.By using the speaker’s voice data,this method obtains the Gaussian Mixed Model（GMM）,and extracts the common overlapping part between the speakers and establishes the common overlap model and non-overlap model.Based on the two models,this method finishes the feature selection of test speech to calculate the similarity in all non-overlapping speaker models,and makes decisions based on the principle of similarity maximizing.Experimental results show that this method is robust,and makes system identification error rate low.

关 键 词：家用机器人 高斯混合模型 特征选择 共性特征 短时说话人 短时语音 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]