基于经验模态分解结合傅氏变换与Wigner分布的Mel频率倒谱系数提取  

Mel Frequency Cepstrum Coefficient Extraction Method Based on Empirical Mode Decomposition and Combined Spectrum of Fourier Transform and Wigner Distribution

作　　者：曾以成[1] 陈雨莺 毛燕湖 谢小娟[1] 

机构地区：[1]湘潭大学物理与光电工程学院,湖南湘潭411105

出　　处：《湘潭大学自然科学学报》2015年第2期20-26,共7页Natural Science Journal of Xiangtan University

摘　　要：根据语音信号的非平稳特点,用经验模态分解方法把语音信号分解成一系列固有模态函数(Intrinsic Mode Function,IMF),一个IMF只含有语音信号的一部分信息,不同IMF分量携带的特征信息不同,对这些IMFs进行加权处理,得到新的语音,再对其进行后续处理.Wigner-Ville分布能精确地定位信号的时频结构,而传统傅氏变换不能反映信号的瞬时变化情况,但多分量信号的Wigner-Ville分布受困于交叉项的干扰,因此利用Wigner-Ville分布的优点,采用Wigner-Ville谱与傅氏谱结合来代替单独的傅氏谱作为每帧的特征,进行Mel频率倒谱系数(Mel Frequency Cepstrum Coefficient,MFCC)提取.实验表明,经改进后的MFCC参数较传统的MFCC参数应用于说话人识别系统,识别率有较大提升,且鲁棒性较好.Speech signal has the non-stationary and nonlinear characteristics,and is decomposed into a number of intrinsic mode functions by applying Empirical Mode Decomposition method.Each IMF contains only part of the information of the speech signal,and different characteristic information carried by different IMF component.Then these IMFs are weighted to get a new speech signal for further processing.WignerVille Distribution can accurately reflect the time-frequency structure of the signal.On the contrary,the Fourier transform can not reflect the instantaneous change of signal.But Wigner-Ville Distribution trapped in cross-term interference by multi-component signals generated.Take advantage of Wigner-Ville Distribution,and using Wigner-Ville spectrum and Fourier spectrum combine to replace Fourier spectrum as the characteristic of each frame for extracting Mel Frequency Cepstrum Coefficient（MFCC）.Experiments show that in speaker recognition system,compared with the classical MFCC parameter,the improved MFCC parameter in this article provides a higher accuracy and better robustness.

关 键 词：经验模态分解 Wigner-Ville谱 傅氏变换 MEL频率倒谱系数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]