一种SVM多类分类算法用于抗噪语音识别  

A multi-class classification method for Support Vector Machine applied to a noise-robust speech recognition

作　　者：侯雪梅[1] 

机构地区：[1]西安邮电学院信息与控制系,陕西西安710121

出　　处：《西安邮电学院学报》2009年第5期100-102,135,共4页Journal of Xi'an Institute of Posts and Telecommunications

基　　金：西安邮电学院中青年教师项目科研基金(110-0417)

摘　　要：为提高机器学习的推广能力,解决语音识别系统在噪声环境中识别率变差等问题,采用改进的MFCC语音特征参数,用支持向量机(SVM)作为语音识别系统的识别网络,对SVM多类分类问题采用"一对一"分类算法,实现了一个汉语孤立词非特定人中等词汇量的抗噪语音识别系统。实验结果表明,SVM线性核函数和多项式核函数具有较好分类结果;当工作在不同信噪比情况下,SVM语音识别系统有较高的识别率,训练时间也能大为缩减,具有较的好鲁棒性。To improve the generalization ability of the machine learning and solve the problem that recc^ition rates of the speech recognition system become worse in the noisy environment, this paper adopted the improved MFCC speech characters, and had the support vector machine as the recognition network for speech recognition system, utilized a one- against- one method for multi- class support vector machine to realize a noise- robust speech recognition system for Chinese isolated words of non- specific person and middle glossary quantity. Experiments showed that the linear kernels and polynomial kernels had better classification results; the SVM speech recognition system had higher correct recognition rates in different SNRs, and was of shorter training time and much better robustness.

关 键 词：支持向量机 核函数 多类分类算法 语音识别 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]