一种改进的听觉特征参数应用于说话人识别  

Improved acoustic characteristic parameter applied to speaker recognition

作　　者：熊冰峰 曾以成[1] 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：《计算机应用》2016年第A01期82-85,共4页journal of Computer Applications

摘　　要：针对主流的语音特征参数梅尔频率倒谱系数(MFCC)和伽马通倒谱系数(GFCC)作为说话人个性特征,在低信噪比的环境下,说话人识别率下降较快的问题,提出一种改进的听觉特征参数应用于说话人识别。在语音信号预处理过程中,提出一种汉明自卷积窗代替经典窗函数,抑制频谱泄露,提高信号的信噪比;从模拟人耳基底膜特性的角度出发,为了更好地模拟基底膜滤波非对称特性,采用全极点Gammatone滤波器提取特征,得到改进的听觉特征参数。在高斯混合模型识别系统中进行仿真实验,实验结果表明,改进的听觉特征参数应用于说话人识别系统,识别正确率都优于线性预测倒谱系数(LPCC)、梅尔频率倒谱系数(MFCC)和Gammatone滤波器倒谱系数(GFCC),而且在信噪比低的环境中,系统仍然有较高的识别率。Focusing on the issue that mainstream speech feature parameters of Mel Frequency Cepstrum Coefficient（ MFCC） and Gammatone Frequency Cepstrum Coefficient（ GFCC） have the poor performance of speaker recognition system under low SNR.An improved auditory parameters applied to speaker recognition was proposed.Hamming self-convolution windows was put forward instead of classic windows in preprocessing to inhibit signal spectrum leakage.Considering that the human basilar membrane,All Pole Gammatone Filter（ APGF） was used to extract characteristic parameter,which get a better simulation of asymmetry filter of Basilar Membrane motion with a simpler parametric modeling.Improved auditory parameters was conducted in Gaussian mixture model recognition system and the result of experiments shows that the improved auditory parameters applied to speaker recognition system has better performance than LPCC,MFCC and GFCC,even in low SNR environment,the system still has high recognition rate.

关 键 词：说话人识别 频谱泄露 汉明自卷积窗 全极点Gammatone滤波器 高斯混合模型 

分 类 号：TN912[电子电信—通信与信息系统]