APP下载

基于辅助函数独立分量分析的频域声回波消除

2023-07-06吴礼福王雷孙芯年孙帅恒

南京信息工程大学学报 2023年3期

吴礼福 王雷 孙芯年 孙帅恒

摘要 传统的声回波消除(Acoustic Echo Cancellation,AEC)方法使用双端通话检测器判断单、双端通话场景,性能受限.盲源分离(Blind Source Separation,BSS)信号模型是一个远端和近端信号并存的全双工模型,因此基于BSS的AEC无需双端通话检测器.本文采用基于辅助函数的独立分量分析(Auxiliary function based Independent Component Analysis,Aux-ICA)算法在频域上实现声回波消除,以最小化互信息为目标函数,借助辅助函数技术进行优化.仿真实验结果表明,在连续的双端通话场景中,该方法具有较低的计算复杂度和较好的回波消除性能.关键词 回波消除;辅助函数;独立分量分析;盲源分离;双端通话

中图分类号TN912

文献标志码A

0 引言

在网络会议、免提通话等应用中,都不同程度地存在声回波问题.回波的存在影响通信质量,严重时会使通信系统不能正常工作.因此,必须采取有效措施来抑制回波,消除其影响.回波消除是通常采用的一种方法,其基本思想是估计出回波路径,得出回波信号的估计,从传声器信号中减去该估计信号,实现回波消除.

自适应滤波[1]是声回波消除的常用方法之一.归一化最小均方(Normalized Least Mean Square,NLMS)算法[2-3]是回波消除的典型算法,该算法通过梯度下降法使估计的回波与麦克风信号之间的均方误差最小.为了防止滤波器发散,需要额外使用双端通话检测器(Double-Talk Detector,DTD)[4]或自适应步长策略[5]来减缓或停止双端通话时自适应滤波器的调整.递归最小二乘法(Recursive Least Square,RLS)[6]也是一种AEC算法,与NLMS算法相比,RLS算法具有更快的收敛速度,但其计算复杂度也更高.Speex MDF[7]是一种广泛使用的自适应滤波回声消除算法,它以NLMS算法为基础,用频域多延时(Multi Delay block Frequency domain,MDF)滤波算法实现,推导出最优步长估计,其优点是滤波器系数基于块更新.

前述的AEC方法存在一定的不足.基于梯度下降的方法存在收敛速度与稳定性之间的平衡问题[8].尽管DTD和自适应步长策略在单向通话和偶尔发生的双端通话场景中都能很好地工作,但在连续双端通话场景中,近端信号总是存在,它们的性能可能会下降[9].盲源分离[10-11]是一种从观测到的混合信号中分离出期望信号来实现信号分离或增强的技术.独立分量分析(Independent Component Analysis,ICA)[12]和独立矢量分析(Independent Vector Analysis,IVA)[13]是典型的BSS技术.AEC可以被认为是一个半盲源分离问题,其目标是从传声器(麦克风)信号中分离出回波和近端信号.

近年来,基于深度学习(Deep Learning)[14-15]的回波消除方法虽然展示了很好的性能,但是这种数据驱动方法主要有两个不足:一是需要足够的数据进行训练,目前虽然有一些开源音频数据库,但这些数据库通常不足以建立鲁棒的神经网络;二是深度神经网络的参数无法解释,这对于希望从自己的需求出发来操纵和调整回波消除系统性能的工程师或实际用户来说是无法接受的.

与传统的AEC算法相比,由于BSS信号模型是一个远端和近端信号并存的全双工模型,所以基于BSS的AEC算法在连续双端通话场景中具有更好的回波消除能力.同时,Speex MDF算法的优异性能表明频域实现AEC具有一定的优势.因而本文采用基于辅助函数的独立分量分析在频域实现声回波消除,在全双工特性的基础上,利用辅助函数技术,避免了显式步长参数选择,降低了算法的计算复杂度.

1 问题描述

1.1 信号模型

1.2 BSS模型

2 Aux-ICA算法

2.1 算法推導

2.2 讨论

Aux-ICA AEC的目标函数是通过最小化互信息得到的,互信息由KL散度(Kullback-Leibler divergence)测量[18],并由辅助函数技术进行优化.在ICA模型中,近端信号被明确地建模为一个独立分量,ICA中的非线性参数β作为加权值.非线性参数β的使用,提高了语音的分离性能.又因为BSS信号模型是远端和近端信号共存的全双工模型,所以Aux-ICA AEC在连续双端通话场景中具有良好的回波消除能力.由于式(21)包含矩阵求逆,计算量较大,并不适合在线应用,可以使用QRD-RLS(QR Decomposition-RLS)算法[19]降低计算复杂度.

在频域进行信号处理时,为防止由于第1帧的回波路径为零矩阵而在信号前端产生较大误差,仿真中需对麦克风信号的第1帧进行预处理,即对第1帧的所有点按照本文算法进行迭代,使得第1帧的回波路径为非全零矩阵.其余帧再根据第1帧进行迭代.Aux-ICA AEC算法消除回波的流程如表1所示.

3 仿真实验

3.1 实验环境

3.2 结果和讨论

4 结论

本文研究了一种基于辅助函数的ICA算法,在频域上实现声回波消除.在全双工特性的基础上,利用辅助函数技术,可以省略显式步长参数选择和双端通话检测器,降低了算法的计算复杂度.仿真验证了该方法具有更低的计算复杂度以及更好的回波消除性能.

参考文献 References

[1]张琦,王霞,王磊,等.自适应回波抵消中变步长NLMS算法[J].数据采集与处理,2013,28(1):64-68ZHANG Qi,WANG Xia,WANG Lei,et al.Variable step-size NLMS algorithms in echo cancellation[J].Journal of Data Acquisition & Processing,2013,28(1):64-68

[2] Morgan D R,Kratzer S G.On a class of computationally efficient,rapidly converging,generalized NLMS algorithms[J].IEEE Signal Processing Letters,1996,3(8):245-247

[3] Albu I,Anghel C,Paleologu C.Adaptive filtering in acoustic echo cancellation systems:a practical overview[C]//2017 9th International Conference on Electronics,Computers and Artificial Intelligence (ECAI).June 29-July 1,2017,Targoviste,Romania.IEEE,2017:1-6

[4] Valin J M.On adjusting the learning rate in frequency domain echo cancellation with double-talk[J].IEEE Transactions on Audio,Speech,and Language Processing,2007,15(3):1030-1034

[5] Yang J.Multilayer adaptation based complex echo cancellation and voice enhancement[C]//2018 IEEE International Conference on Acoustics,Speech and Signal Processing.April 15-20,2018,Calgary,AB,Canada.IEEE,2018:2131-2135

[6] 丁鋒.辅助模型辨识方法(5):最小二乘辨识[J].南京信息工程大学学报(自然科学版),2016,8(5):385-403DING Feng.Auxiliary model based identification methods.Part E:least squares identification[J].Journal of Nanjing University of Information Science & Technology (Natural Science Edition),2016,8(5):385-403

[7] 杨果,王亚洁,王晶.Speex编码器中回声消除算法的分析与评估[J].电声技术,2013,37(9):52-55YANG Guo,WANG Yajie,WANG Jing.Analysis and evaluation of echo cancellation algorithm in speex codec[J].Audio Engineering,2013,37(9):52-55

[8] Ono N,Miyabe S.Auxiliary-function-based independent component analysis for super-Gaussian sources[C]//Latent Variable Analysis and Signal Separation,2010:165-172

[9] Gunther J.Learning echo paths during continuous double-talk using semi-blind source separation[J].IEEE Transactions on Audio,Speech,and Language Processing,2012,20(2):646-660

[10] Ikram M Z.Blind source separation and acoustic echo cancellation:a unified framework[C]//2012 IEEE International Conference on Acoustics,Speech and Signal Processing.March 25-30,2012,Kyoto,Japan.IEEE,2012:1701-1704

[11] 张艳萍,李杰.回波对消中盲源分离算法的研究与仿真[J].南京信息工程大学学报(自然科学版),2013,5(6):539-543ZHANG Yanping,LI Jie.Optimization and simulation of blind source separation in acoustic echo cancellation[J].Journal of Nanjing University of Information Science & Technology (Natural Science Edition),2013,5(6):539-543

[12] 石怀涛,周乾,王雨桐,等.基于相对变换的ICA故障检测方法[J].电子测量与仪器学报,2017,31(7):1040-1046SHI Huaitao,ZHOU Qian,WANG Yutong,et al.Fault detection method based on relative-transformation ICA[J].Journal of Electronic Measurement and Instrumentation,2017,31(7):1040-1046

[13] Scheibler R,Ono N.Independent vector analysis with more microphones than sources[C]//2019 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics.October 20-23,2019,New Paltz,NY,USA.IEEE,2019:185-189

[14] Lei Q H,Chen H,Hou J F,et al.Deep neural network based regression approach for acoustic echo cancellation[C]//Proceedings of the 2019 4th International Conference on Multimedia Systems and Signal Processing.2019:94-98

[15] 周曉彦,王珂,李凌燕.基于深度学习的目标检测算法综述[J].电子测量技术,2017,40(11):89-93ZHOU Xiaoyan,WANG Ke,LI Lingyan.Review of object detection based on deep learning[J].Electronic Measurement Technology,2017,40(11):89-93

[16] Taniguchi T,Ono N,Kawamura A,et al.An auxiliary-function approach to online independent vector analysis for real-time blind source separation[C]//2014 4th Joint Workshop on Hands-free Speech Communication and Microphone Arrays (HSCMA).May 12-14,2014,Villers-les-Nancy,France.IEEE,2014:107-111

[17] Ono N.Auxiliary-function-based independent vector analysis with power of vector-norm type weighting functions[C]//Proceedings of the 2012 Asia Pacific Signal and Information Processing Association Annual Summit and Conference.December 3-6,2012,Hollywood,CA,USA.IEEE,2012:1-4

[18] 魏本征,甘洁,尹义龙.基于边缘特征点互信息熵的医学图像配准方法[J].数据采集与处理,2018,33(2):248-258WEI Benzheng,GAN Jie,YIN Yilong.Medical image registration based on mutual information entropy combined with edge correlation feature[J].Journal of Data Acquisition and Processing,2018,33(2):248-258

[19] 姜波,杨军,张尔扬.稀疏自适应Volterra滤波的QRD-RLS算法[J].信号处理,2008,24(4):595-599JIANG Bo,YANG Jun,ZHANG Eryang.QRD-RLS algorithm for sparse adaptive Volterra filtering[J].Signal Processing,2008,24(4):595-599

[20] Allen J B,Berkley D A.Image method for efficiently simulating small-room acoustics[J].The Journal of the Acoustical Society of America,1979,65(4):943-950

[21] Loizou P C.Speech enhancement:theory and practice[M].Florida:CRC Press,2007:492-495

Frequency domain acoustic echo cancellation using auxiliaryfunction based independent component analysis

WU Lifu WANG Lei SUN Xinnian SUN Shuaiheng

1School of Electronics & Information Engineering,Nanjing University of Information Science & Technology,Nanjing 210044

2Collaborative Innovation Center of Atmospheric Environment and Equipment Technology,

Nanjing University of Information Science & Technology,Nanjing 210044

AbstractThe performance of traditional Acoustic Echo Cancellation (AEC) is restricted due to the double-talk detector it used to determine the double-talk and single-talk scenarios.While Blind Source Separation (BSS) signal model is a full duplex model with both far-end and near-end signals,thus the BSS-based AEC does not need the double-talk detector.This paper adopts Auxiliary function based Independent Component Analysis (Aux-ICA) algorithm to realize acoustic echo cancellation in frequency domain,in which the object function is minimizing the mutual information,and the auxiliary function technique is used for optimization.Simulation results show that this method has lower computational complexity and better performance in acoustic echo cancellation under continuous double-talk scenarios.

Key words echo cancellation;auxiliary function;independent component analysis (ICA);blind source separation;double-talk