Abstract: The optimal selection of the speech frames is important and necessary to generate the speaker template of the speaker recognition system since the number of the frames is too large.The existing general selection procedures based on the large number of enumeration and many times iteration,are usually complicated and time-consuming,and the result generated by these methods is not always optimal.The Support Vector Machines(SVM) based on ethe Statistical Learning Theory can solve this problem.An improved SVM,the Least Square Support Vector Machines(LS-SVM) is discussed in this paper.The experimental results demonstrate that the LS-SVM-based speaker recognition is less computational complexity and more effient than the SVM-based speaker recognition.Then it has high adaptability for the speaker recognition.

Key words: kernel function, linear predictive coding, speaker recognition, least square support vector machines

摘要： 说话人识别系统在说话人模板的建立过程中由于说话人的语音帧数量太多,往往要进行筛选,通常这种选择是一种基于枚举的大量反复的提取过程,复杂费时而结果往往并不是最优的。而基于统计学习理论的支持向量机(SVM) 方法正好克服了这方面的不足。本文讨论了一种改进的SVM即最小二乘向量机(LS-SVM)的方法进行说话人识别研究。研究表明，基于LSSVM的说话人识别比传统的SVM说话人识别计算复杂度小、效率更高、对说话人识别有很强的适应性。

关键词: 核函数, 线性预测, 说话人识别, 最小二乘向量机