PCA-VQ融合降维的SMO-SVM说话人识别研究  

Research on SMO-SVM Speaker Recognition Based on PCA-VQ Fusion Dimensionality Reduction

作　　者：席青云 孙同日 陶佰睿[3] 杨文博 苗凤娟[3] XI Qingyun;SUN Tongri;TAO Bairui;YANG Wenbo;MIAO Fengjuan(XingAn League Branch,Inner Mongolia Radio and Television University,XingAn League,Inner Mongolia 137400,China;Computer Science and Information Engineering,Heihe University,Heihe Heilongjiang 164300,China;College of Communications and Electronics Engineering,Qiqihar University,Qiqihar Heilongjiang 161006,China)

机构地区：[1]内蒙古广播电视大学兴安盟分校,内蒙古兴安盟137400 [2]黑河学院计算机与信息工程学院,黑龙江黑河164300 [3]齐齐哈尔大学通信与电子工程学院,黑龙江齐齐哈尔161006

出　　处：《传感技术学报》2023年第2期275-279,共5页Chinese Journal of Sensors and Actuators

基　　金：教育部“科教融创”职业教育改革创新课题(HBKC217107);黑龙江省教育厅基本科研业务专项(145209804)。

摘　　要：针对说话人语音原始梅尔频率倒谱系数(MFCC)特征参数维数较高造成的模型计算效率低以及不稳定的问题,基于序列最小优化(SMO)高效算法求解支持向量机(SVM)基本型的对偶问题,开展主成分分析-矢量量化(PCA-VQ)融合降维的SMO-SVM说话人识别算法研究。改进后的算法在MATLAB平台上仿真通过。仿真结果表明:通过PCA-VQ融合算法对MFCC特征参数进行优化降维后,SMO-SVM说话人识别模型的正确率提高3.77%,训练时间节省1.24 s,具有较好推广应用价值。Aiming at the problem of low computational efficiency and instability of the model caused by high dimension of the original Mel Frequency Cepstrum Coefficient(MFCC)characteristic parameters of speaker’s speech,the dual problem of the basic form of Support Vector Machine(SVM)is solved based on Sequence Minimum Optimization(SMO)efficient algorithm,and SMO-SVM speaker recognition algorithm with principal component analysis-vector quantization(PCA-VQ)fusion dimensionality reduction is developed.The improved algorithm is simulated on the MATLAB platform.The simulation results show that after optimizing the MFCC feature parameters and reducing the dimension through the PCA-VQ fusion algorithm,the accuracy of SMO-SVM speaker recognition model is improved by 3.77%and the training time is saved by 1.24 s,which has good popularization and application value.

关 键 词：说话人识别 主成分分析 矢量量化 序列最小优化 支持向量机 

分 类 号：TN912.3[电子电信—通信与信息系统] TP183[电子电信—信息与通信工程]