基于特征加权的FSVM在低信噪比语音识别中的应用  

Fuzzy support vector machines based on feature weighting in low SNR speech recognition application

作　　者：刘红芬[1] 张雪英[1] 刘晓峰[2] 黄丽霞[1] 王子中[1] 

机构地区：[1]太原理工大学信息工程学院,太原030024 [2]太原理工大学理学院,太原030024

出　　处：《太原理工大学学报》2014年第6期764-768,共5页Journal of Taiyuan University of Technology

基　　金：国家自然科学基金:基于认知机理的情感语音识别基础研究(61371193);山西省青年科技研究基金(2013021016-1);山西省2012年留学回国人员科技活动择优资助

摘　　要：模糊支持向量机(FSVM)赋予每个样本一个模糊隶属度,优化了最优分类面,具有好的抗噪性。为提高低信噪比条件下的语音识别准确率,减小传统支持向量机(SVM)中噪声样本对分类精度造成的影响,首先将模糊支持向量机(FSVM)应用于语音识别系统中的识别网络。在此模型的基础上,考虑到语音样本各维特征的影响,计算各维特征的信息增益作为对分类的重要程度,依据此重要程度确定对各维特征的加权系数,构造加权核函数。为了减小一对一分类模式下的时间复杂度,引入了三叉决策树策略。在仿真实验中,采用韩语语料库的耳蜗滤波器倒谱系数(CFCC)作为实验数据,实验结果表明,在低信噪比语音识别中,特征加权模糊支持向量机(FWFSVM)比传统的SVM具有更好的鲁棒性。. Fuzzy support vector machine （FSVM） has good noise immunity by giving eachsample a fuzzy membership to optimize the optimal classification surface. To improve the recogni-tion accuracy in the low SNR speech recognition system, and to reduce the influence of noises oroutliers on the classification accuracy in the traditional support vector machines, FSVM is proposed in the recognition network of speech recognition system. Based on this, to take the impor-tance of each feature into account in the classification task, and by computing the relative impor-tance（weight）of each feature, the feature weighted inner product is constructed. Triple decisiontree is introduced to reduce the computational complexity in one-versus-one method. During sim-ulation experiment, cochlear feature cepstral coefficients（CFCC） features are adopted, the resultof the experiment indicates that Feature Weighted FSVM possesses better robustness than tradi-tional SVM in low SNRs.

关 键 词：模糊支持向量机 三叉决策树 加权 低信噪比 鲁棒性 

分 类 号：TN912[电子电信—通信与信息系统]