说话人识别算法鲁棒性研究  

Research into the Robustness of Algorithm for Speaker Recognition

作　　者：冯月芹[1] 郝雯超 陈义 王蕾 李春光 

机构地区：[1]南京工程学院通信工程学院,江苏南京211167

出　　处：《南京工程学院学报（自然科学版）》2017年第3期60-66,共7页Journal of Nanjing Institute of Technology(Natural Science Edition)

基　　金：南京工程学院创新基金面上项目(CKJC201505)

摘　　要：由于噪声干扰问题,说话人识别算法的效率受到很大影响.为此,在美尔倒谱系数特征提取的基础上,采用半升正弦函数对特征进行修正.采用特征规整、特征弯折和特征映射三种方法对特征进行校对.为提高说话人识别率以及模型的鲁棒性,提出混合BP神经网络与混合高斯模型的方法,将高斯混合模型的概率输出作为神经网络输入,从而获取说话人间的交互信息.试验结果显示,算法的识别率高,抗噪性好.The efficiency of speaker recognition algorithmis greatly affected by noise disturbance. the characteristics are proposed to be corrected by using half raised-sine function （ HRSF） based on the extraction of Melfrequency cepstrum coefficients （ MFCC） . Three methods, i. e. , feature normalization, feature warping, and feature mapping, are then adopted to further correct the characteristics. Meanwhile, to improve speaker recognition rate androbustness of the model, this paper proposes a combination of BP neural network （ NN） and Gaussian mixture mo（GMM）, so that GMM probability is put into the neural network and the information of interaction between differentspeakers is obtained. The results showa higher recognition rate and a better noise immunity.

关 键 词：说话人识别 美尔频率倒谱系数 高斯混合模型 神经网络 

分 类 号：TN912.3[电子电信—通信与信息系统]