基于双耳相干函数的鲁棒时延差估计与混响抑制算法  

A robust time-delay estimation and dereverberation algorithm based on the coherence function

作　　者：方义 陈友元[2] 牟宏宇[2] 冯海泓[2] FANG Yi;CHEN Youyuan;MOU Hongyu;FENG Haihong(The Institute of Acoustics, University of Chinese Academy of Sciences, Beijing 100190, China;Shanghai Acoustics Laboratory, Chinese Academy of Sciences, Shanghai 200815, China)

机构地区：[1]中国科学院大学声学研究所,北京100190 [2]中科院声学研究所东海研究站,上海200815

出　　处：《清华大学学报（自然科学版）》2018年第5期516-522,共7页Journal of Tsinghua University(Science and Technology)

基　　金：国家自然科学基金资助项目(11474309)

摘　　要：传统的基于相关峰的广义互相关算法在混响环境下性能急剧下降,尽管一些优先效应模型被提出以改善其性能,但是这些模型计算复杂且对阈值选取很敏感。该文首先通过协方差矩阵的特征值来分别更新语音的相干函数和噪声的相干函数,随后将语音的相干函数与理想相干函数匹配,用于时延差估计。估计出的时延差和噪声的相干函数用于相干与散射信号能量比值（coherent-to-diffuse power ratio,CDR）的估计,最后利用实时估计出来的CDR值进行混响抑制。实验结果表明：该方法的定位误差明显低于传统方法,且混响抑制后的主观语音质量评估（perceptual evaluation of speech quality,PESQ）分数高于对比算法。The performance of traditional cross-correlation based time-delay estimation methods is sharply degraded in reverberation environments.Precedence effect models have been proposed with cross-correlation functions, but these models are quite parameter-sensitive and the front-end processes are very complex.This paper describes a method that first updates a function of the speech and noise based on the eigenvalues of the covariance matrix.Then,a coherence function of the speech is matched to the ideal coherence function for the time-delay estimate.Then,the estimated time delay and the noise coherence function are applied to the coherent-to-diffuse power ratio（CDR）estimator for reverberation suppression.Tests show that this scheme has higher localization accuracy than traditional methods and achieves higher PESQ（perceptual evaluation of speech quality） scores than other CDR estimators.

关 键 词：优先效应模型 时延差估计 混响抑制 相干函数 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置] TN912.34[自动化与计算机技术—控制科学与工程]