摘要
针对现有说话人识别系统识别率不高,鲁棒性能差的缺点,提出了一种基于超向量的核函数构造方法。通过对超向量进行KL散度变换和L2线性内积变换,分别得到KL散度线性核函数、KL散度非线性核函数以及L2内积核函数。实验结果表明,将这三种核函数分别应用于支持向量机的说话人识别系统,可以得到优于常规核函数的识别性能。
In order to improve the recognition ratio and weak robustness of speaker recognition system, a new SVM-classifier based on supervector kernel function is proposed in this paper. These new kernels, KL divergence linear kernel function, KL nonlinear kernel function and L2 inner product kernel function based on supervetor, are generated. Then these new kernel functions are applied into speaker recognition system and good experiment results are achieved.
出处
《微计算机信息》
2009年第7期254-256,共3页
Control & Automation
关键词
说话人识别
KL散度
GMM超向量
L^2内积核函数
核函数
speaker recognition
KL divergence
GMM supervector
L^2inner product kernel function
Kernel Function
