偏度最大化多通道逆滤波语声去混响研究  

Maximum skewness-based multichannel inverse filtering for speech dereverberation

作　　者：郭颖 彭任华 郑成诗[1] 李晓东[1] GUO Ying;PENG Renhua;ZHENG Chengshi;LI Xiaodong(Key Laboratory of Noise and Vibration Research,Institute of Acoustics,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院噪声与振动重点实验室(声学研究所),北京100190 [2]中国科学院大学,北京100049

出　　处：《应用声学》2019年第1期58-67,共10页Journal of Applied Acoustics

基　　金：国家自然科学基金项目(61571435)

摘　　要：房间混响会降低语声质量和语声可懂度。高阶统计量是衡量非高斯性的重要参量,基于语声非高斯特性可实现语声去混响。该文提出一种基于高阶统计量的多通道语声去混响方法,该方法首次用多通道语声信号线性预测残差的三阶统计量偏度构造代价函数,以去混响重建信号线性预测残差的偏度最大化为目标自适应地更新逆滤波器,同时引入通道逆滤波和语声产生系统的联合估计。实验结果表明,该方法相较于已有的基于线性预测残差四阶统计量峰度的方法具有更好的去混响效果,且对噪声具有更强的鲁棒性。Room reverberation often leads to the reduction of speech quality and speech intelligibility.Speech dereverberation can be achieved by using non-Gaussian property of speech,where higher order statistics (HOS) are typical measurements.This paper presents a method based on HOS for multichannel speech dereverberation.The cost function is constructed using the third-order statistics,namely skewness,of multichannel speech signal linear prediction residuals,and then update the inverse filter adaptively by maximizing the skewness of the linear prediction residuals of the reconstructed speech signal.Meanwhile,we introduce the joint estimation of the channel’s inverse filter and the speech production system.Experimental results show that the proposed method is superior to the method based on forth-order statistics,i.e.kurtosis,in terms of dereverberation and robustness to the noise.

关 键 词：高阶统计量 偏度 线性预测 房间脉冲响应 逆滤波 

分 类 号：TN912.35[电子电信—通信与信息系统]