一种鲁棒性听觉特征的说话人确认系统  

A Speaker Verification System with Robust Auditory Feature

作　　者：文思进 高勇[1] WEN Sijin;GAO Yong(College of Electronics and Information Engineering, Sichuan University, Chengdu 610065, China)

机构地区：[1]四川大学电子信息学院

出　　处：《无线电工程》2019年第7期606-610,共5页Radio Engineering

基　　金：四川大学科研基金资助项目(0020505501743)

摘　　要：针对噪声环境下说话人确认系统性能急剧下降问题,根据人耳听觉感知特性,利用Gammachirp滤波器组来模拟人耳耳蜗听觉模型,提出了一种鲁棒性听觉特征参数(GammachirpFeatureCoefficient,GCFC)的提取方法。在高斯混合模型-通用背景模型(GaussianMixtureModel-UniversalBackground,GMM-UBM)下进行仿真实验,研究了不同噪声环境下系统的噪声鲁棒性和适应性。实验结果表明,在说话人确认系统中,新提取的听觉特征参数在噪声鲁棒性、噪声适应性和系统整体确认性能上均优于梅尔倒谱系数和基于Gammatone滤波器的听觉特征参数。According to human auditory perception characteristics,an robust Gammachirp feature coefficient(GCFC) extraction method is introduced for the matter of sharp performance degradation in speaker verification system under the noisy environment.This method is based on Gammachirp filter banks which can simulate the auditory model of cochlea.The simulation experiments are conducted in Gaussian Mixture Model-Universal Background Model(GMM-UBM) subsequently,including the noise robustness and adaptability experiments of the system under different noisy environments.The experiment results show that,in speaker verification system,the newly extracted GCFC feature parameters are better than Mel cepstral coefficients and GFCC feature parameters based on Gammatone filter banks in the noise robustness,noise adaptability and overall verification performance.

关 键 词：说话人确认 Gammachirp滤波器 高斯混合模型-通用背景模型 鲁棒性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]