基于Relief+SVM的语音信号特征提取及其识别  

The Feature Extraction and Its Recognition of Speech Signal Based on Relief and SVM

作　　者：苏盈盈[1,2] 刘兴华[1] 葛继科[1] 李太福[1] 文峰[1] 

机构地区：[1]重庆科技学院电气与信息工程学院,重庆401331 [2]重庆大学自动化学院,重庆400044

出　　处：《重庆科技学院学报（自然科学版）》2013年第5期122-125,共4页Journal of Chongqing University of Science and Technology：Natural Sciences Edition

基　　金：重庆市自然科学基金项目(cstc2012jjA40026);重庆科技学院校内科研基金项目(CK2011Z01)

摘　　要：针对语音信号的原始波形不能直接用于识别的问题,提出利用Relief算法对语音信号进行特征提取,在降低输入特征维数的同时,保留原有信号的特征。再利用结构风险最小化的支持向量分类器进行语音信号的识别。最后通过实例分析,对民歌、古筝2种不同音乐进行特征提取后,从原有的24维特征降为19维,分类器的正确率达到95.6%,而且算法的运行时间明显缩短。仿真结果表明该方法在语音信号特征选择,提高算法运行速度方面有明显效果。The original waveform of speech signals cannot be directly used for identification of problems.According to it,it is studied based on Relief algorithm for feature selection.After the extraction of speech signal,the information characteristics of original signal is retained,and then SVM（support vector classifier） with structure risk minimization is used to identify the speech signal.Finally throughout the instance analysis,two kinds of different music are extracted from the initial 24 feature dimension reduction to 19 dimension,which are the folk song and the guzheng.Meanwhile,the classifier accuracy reaches 95.6%,and the algorithm running time is significantly shortened.The simulation results show that the method is not only effective in the feature selection of speech signal,but also the obvious effect to improve the running speed of algorithm.

关 键 词：RELIEF 支持向量分类器 语音信号 特征选择 

分 类 号：TN912.3[电子电信—通信与信息系统]