基于支持向量机的变异语音分类研究  

Stressed speech classification based on support vector machines

作　　者：王欢良[1] 韩纪庆[1] 张磊[1] 

机构地区：[1]哈尔滨工业大学计算机科学与工程系,黑龙江哈尔滨150001

出　　处：《哈尔滨工业大学学报》2003年第4期389-393,共5页Journal of Harbin Institute of Technology

基　　金：国家自然科学基金资助项目 (60 0 85 0 0 1) ;教育部留学回国人员科研启动基金项目资助

摘　　要：变异语音的训练样本有限 ,采用传统的分类方法进行分类 ,效果不够理想 ,而支持向量机方法在有限样本情况下可以保持很好的分类推广能力 .采用支持向量机方法进行变异语音分类 ,提取基频和TEO基频作为变异语音分类的特征 ,讨论了样本预处理和参数选择等问题 .提出了直接截取和DTW规正两种方法来解决语音样本特征向量长度不一致的问题 .基于TEO基频特征 ,采用指数径向基函数 (ERBF)内核 ,对应力(G -force)影响下的变异语音进行分类 ,分类正确率可达到 99 2 % ,比传统的贝叶斯分类器和HMM分类器 ,分类性能分别平均提高了 12 6 %和 6 0 % .实验结果表明 ,采用支持向量机方法进行变异语音分类是可行的 .The support vector machines (SVMs) method is applied to the stressed speech classification to extract pitch frequency and Teager energy operator (TEO) pitch frequency as the features of stressed speech classification. The preprocessing of samples and the selection of parameters are discussed. For the lengths of feature vectors of speech samples are different, the direct cutting and the dynamic time warping (DTW) regulation are put forward to solve the problem. When TEO pitch frequency feature and exponential radial basis function (ERBF) kernel function are used to classify the stressed speech under G-force, the correct classification probability can reach 99.2%. In comparison with those of traditional Bayes classifier and hidden Markov model (HMM) classifier, its classification performances are improved by 12.6% and 6.0% respectively. Experimental results indicate that it is feasible to adopt SVMs for the stressed speech classification.

关 键 词：支持向量机 变异语音分类 样本预处理 参数选择 TE0基频 语音信号处理 

分 类 号：TN912.34[电子电信—通信与信息系统]