WCCN聚类序列核函数在话者识别中的应用

A novel WCCN clustering kernel applied in speaker recognition

针对说话人确认系统中GMM超向量建模计算复杂度高以及易受信道干扰的问题,提出一种新型的基于Bhattacharyya距离聚类的WCCN序列核函数算法.首先计算话者GMM模型之间的Bhattacharyya距离,根据该Bhattacharyya距离对话者模型进行聚类,得到聚类中心模型;紧接着对聚类中心模型的均值向量进行MAP自适应,进而生成超向量序列核函数;最后采用WCCN平滑归一化技术对序列核函数进行信道补偿,抑制噪音和信道畸变对核函数的影响.将该Bhattacharyya聚类WCCN核函数应用到SVM说话人确认系统,仿真实验结果表明该核函数可以有效地提高系统的识别准确率和识别速度.

A novel WCCN kernel based on Bhattacharyya distance clustering algorithm was proposed in this paper in order to reduce the computation complexity of GMM super-vector,meanwhile the channel interference was removed from speaker verification system.Firstly,the GMM models of speakers were clustered based on Bhattacharyya distance,and clustering center models were obtained.Then super-vector sequence kernel was generated by adapting only mean vectors of these clustering center models.Finally,WCCN was used to restrain the noise and channel distortion effection of this kernel.Our experiment results showed that our new kernel can improve the recognition accuracy and speed.