基于信道补偿的说话人识别算法  

Speaker recognition algorithm based on channel compensation

作　　者：申铉京[1,2] 翟玉杰[1,2] 卢禹彤 王玉[1,2,4] 陈海鹏[1,2] 

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学符号计算与知识工程教育部重点实验室,长春130012 [3]香港理工大学工程学院,香港999077 [4]吉林大学应用技术学院,长春130012

出　　处：《吉林大学学报（工学版）》2016年第3期870-875,共6页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家青年科学基金项目(61305046);吉林省自然科学基金项目(20140101193JC);吉林省青年科学基金项目(20130522117JH)

摘　　要：现有说话人识别算法普遍受信道因素的干扰,为了提高算法的准确率,在特征级利用特征弯折算法对语音特征参数进行处理,在模型级利用因子分析技术对说话人混合高斯模型(GMM)进行信道处理。对端点进行检测后,利用特征弯折算法对语音特征参数梅尔倒谱系数(MFCC)进行处理,去除线性信道和背景噪声的影响,并建立说话人GMM。然后利用因子分析技术拟合说话人特征空间与信道空间的差异,去除信道因子的影响。最后提取高斯超向量并通过支持向量机(SVM)得到识别结果。实验结果证明了信道补偿算法与GMM-SVM相结合能获得更好的识别率,并能保证算法的鲁棒性。Channel interference factor for the identification results is prevalent among the existing speaker recognition algorithm.In order to improve the accuracy of the system,in this paper,feature warping is used to compensate the channel factor of Mel-Frequency Cepstral Coefficient（MFCC）features.Then,factor analysis technique is applied to deal with the channel factors of the speaker＇s Gaussian Mixture Model（GMM）.In the endpoint detection phase of speech of this recognition system,the GMM for speech modeling is built to accurately determine the beginning and end points of the speech segment,and then the features after feature warping are used to establish speaker GMM.Using factor analysis technique to fit the differences between the speaker characteristics space and the channel space,the algorithm removes channel factor from GMM,and then extracts GMM super-vectors as input of the Support Vector Machine（SVM）to obtain recognition results.Experimental results show that the combination of channel compensation technique and SVM can obtain better recognition rate,and ensure the robustness of the system.

关 键 词：计算机应用 说话人识别 支持向量机 混合高斯模型 特征弯折 隐藏因子分析 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]