分类号
TN912.35
密级
UDC 004.93
学位论文编号
重庆邮电大学硕士学位论文
中文题目
声学回声消除进阶问题中的高效自适应滤
波算法研究
英文题目
Study on Efficient Adaptive Filtering
Algorithms for Advanced AEC Problems
学 号
S150101216
姓 名
周俊锴
学位类别
工学硕士
学科专业
信息与通信工程
指导教师
周翊
完成日期
2018 年 3 月 25 日
重庆邮电大学硕士学位论文 摘要
I
摘要
声学回声消除(Acoustic Echo Cancellation, AEC)技术被广泛应用到免提通信中,
它使得通话双方在无需手持终端设备舒适进行通话的过程中免除回声带来的干扰。
由于 AEC 技术应用的场景趋于多样化和复杂化,传统的单通道线性 AEC 已经不
能满足市场需求。立体声声学回声消除(Stereophonic Acoustic Echo Cancellation,
SAEC)和非线性声学回声消除(Nonlinear Acoustic Echo Cancellation, NLAEC)这两
类进阶 AEC 技术应运而生。本文分别从自适应滤波算法的鲁棒性、计算复杂度和
稳态收敛性着手,对应用于 SAEC 的两路自适应滤波算法和应用于 NLAEC 的核函
数自适应滤波算法进行深入研究,主要工作的内容如下:
首先,针对 SAEC 场景中经常出现的双端讲话和回声路径突变情况,本文提
出了一种双通道的双路径频域算法。区别于传统的 ELMS 算法中对每条回声路径
利用一个滤波器跟踪的方法,本文采用双滤波器结构,通过一个持续更新的背景
滤波器和一个非自主更新的前景滤波器同时跟踪一条回声路径,以提高算法在高
干扰情况下的鲁棒性。该算法通过背景滤波器持续跟踪回声路径,并由转移逻辑
控制其在高干扰环境下将抽头系数传递给前景滤波器来输出稳态误差。这不仅解
决了双端讲话时自适应滤波器容易发散的问题,同时还克服了回声路径突变时传
统双讲检测(Double Talk Detection, DTD)误判所造成的滤波器系数停止更新问题。
在此基础上,本文将双路径频域算法移植到 WebRTC 平台的 AEC 模块中,实现并
验证了算法的工程应用。
其次,本文为实现核函数自适应滤波算法在计算复杂度和收敛速度上的优化
配置,提出了一种用于 NLAEC 的改进型核函数自适应滤波切换算法。该算法首先
利用基于快慢包络检测的语音活动检测(Voice Activity Detection, VAD)技术判别语
音段的不同能量,然后采用快包络和慢包络差值运算求出能量阈值,进而通过语
音信号能量的不同对核自适应滤波算法进行切换。核函数自适应滤波切换算法不
仅能够具有传统线性自适应滤波算法不具有的对非线性信道进行估计建模的能力,
而且,通过引入切换机制,其能够很好地平衡收敛速度和运算复杂度之间的关系,
是一种适于工程实现的 NLAEC 解决方案。
重庆邮电大学硕士学位论文 摘要
II
关键词:声学回声消除,双滤波器结构,双讲检测,核自适应滤波
重庆邮电大学硕士学位论文 Abstract
III
Abstract
Acoustic Echo Cancellation (AEC) technology is widely used in hands-free
communication, which makes it possible for both parties to cancel the interference
caused by the echo during the conversation without holding the terminal equipment.
Due to the diversification and complexity of the application of AEC technology, the
traditional single-channel linear AEC becomes unable to meet the market’s demands.
Stereophonic Acoustic Echo Cancellation (SAEC) and Nonlinear Acoustic Echo
Cancellation (NLAEC) are two advanced AEC technologies. This thesis focuses on the
robustness, computational complexity, and steady-state convergence of adaptive
filtering algorithms, which are the cores of the dual-path adaptive filtering algorithm for
SAEC and kernel adaptive filtering algorithm applied to NLAEC. The thesis is
organized as follows:
Firstly, aiming at the double-talk and echo path change that often appear in the
SAEC scenario, this thesis proposed a dual-channel dual-path frequency domain
algorithm. Different from the traditional ELMS algorithm, which utilizes one filter to
track each echo path, a double filter structure is applied in the proposed algorithm to
track an echo path simultaneously through a continuously updated background filter and
a non-auto-updating foreground filter so as to improve the robustness of the algorithm
in high interference environments. The algorithm tracks the echo path continuously
through the background filter, and the transition logic controls it to transfer the tap
coefficients to the foreground filter in a high-interference environment to output the
steady-state error. This method not only solves the problem that the adaptive filter is
easy to diverge during double-talk, but also overcomes the problem of filter coefficients
freezing caused by the misjudgment of traditional double talk detection (DTD) when the
echo path suddenly changes. Next, the proposed dual-path frequency domain algorithm
is transplanted to the AEC module of WebRTC platform, realize and validate the
engineering application of this algorithm.
Secondly, aiming to achieve optimal configuration of the computational
complexity and convergence speed, a switching adaptive kernel adaptive filter
switching algorithm for NLAEC is proposed. The algorithm first utilizes a voice activity
detection (VAD) technique based on the fast and slow envelope detection to