CPSO聚类和KMP相结合的说话人识别方法

Speaker Recognition Method Based on CPSO Clustering and KMP Algorithm

针对目前广泛应用于说话人识别领域的MFCC特征参数包含较少说话人特征信息的问题和SVM分类器选择核函数时受到Mercer准则限制的问题,提出了一种将混沌粒子群算法(CPSO)与核匹配追踪算法(KMP)相结合的说话人识别方法.首先通过CPSO聚类算法将MFCC特征参数进行变换处理,得到精简的MFCC特征参数(SMFCC),然后利用KMP算法对核函数的形式没有任何限制的特性和良好的分类识别性能,对约简后的SMFCC特征参数进行分类训练和识别.仿真实验结果表明,基于CPSO-KMP说话人识别方法相比主流的GMM-UBM方法,在EER性能上相对提高了31%.

MFCC feature parameters and SVM classifier have been widely used in the field of speaker recognition,but each of them has its ow n limitation.MFCC feature parameters contain less speaker characteristics,w hile SVM classifier is restricted by Mercer in kernel function selection.Therefore,this paper proposes a new speaker recognition method,w hich combines chaotic particle sw arm optimization（CPSO） and kernel matching pursuit algorithm（KMP）.First of all,transform the MFCC feature parameters by CPSO clustering algorithm,to get simple MFCC（SMFCC）.Then classify and recognize it by KMP,for KMP has better performance of classification and recognition,w ithout any kernel function selecting restriction.The result show s that,compared w ith mainstream GMM-UBM method,CPSO-KMP speaker recognition method increases 31% in EER performance.