一种改进型的MEL滤波器混合特征参数提取方法研究  

A Modified Mixture Feature Parameter Extraction Method of Mel- Frequency Filter

作　　者：黄锐[1] 陆安江[1] 张正平[1] 

机构地区：[1]贵州大学大数据与信息工程学院,贵州贵阳550025

出　　处：《通信技术》2014年第12期1388-1391,共4页Communications Technology

基　　金：贵州省社发攻关项目(黔科合SY字[2013]3105号)~~

摘　　要：由于在说话人识别中梅尔频率滤波器组结构分布不均匀,在低频区域分布密集而在中心频率、高频率分布稀疏,影响了在中、高频段的MEL倒谱系数(MFCC)的提取,本文提出适用于说话人识别的改进MEL滤波器与Mid Mel滤波器相结合得到两种混合特征参数,用此方式来提高中、高频率特征参数提取的精度,从而提高系统识别率。实验结果显示,在同一环境中,新的混合特征参数识别率与识别性能优于传统的特征参数,且运算量较少。The nonuniform distribution of mel-fiherbank structure in speaker recognition, that is, too inten- sive in low frequency region while too sparse in high frequency and mid-frequency region, would affect the extraction of MFCC in mid-frequency and high frequency. In light of this, the paper proposes a method to extract the two mixed feature parameters by combining Mel-fiherbank and MidMel-filterbank, this method applicable to speaker recognition system could improve the accuracy of medium and high frequency feature parameters extraction and thus upgrade the system recognition rate. Experimental results indicate that the recognition rate and recognition performance of the novel mixed parameters is superior to that of traditional characteristic parameters, and moreover, with less computational complexity.

关 键 词：说话人识别 梅尔频率滤波器 MidMel滤波器 MEL倒谱系数 

分 类 号：TN912.34[电子电信—通信与信息系统]