立體聲回音消除器之去相關性演算法與共極點架構之研究

Abstract

在立體聲回音消除器的運作中，遠端輸入訊號之間的高相關性會造成適應性濾波器無法快速地估計到回音路徑。本篇論文針對兩類能夠增加適應性濾波器收斂速度的方法做討論：降低適應性濾波器輸入訊號之間的相關性與共極點架構的問題。第一部分著重於將互斥最大化係數選擇法與非線性成分加強法做了兩種不同的結合方式(形式1與形式2)，發現此兩種結合法皆能夠進一步帶來收斂速度的改善，其中形式2的收斂速度又優於形式1。第二部分研究共極點架構，藉由降低適應性濾波器使用的階數來提升收斂速度，但是共極點近端等化架構具有近端失真、殘留回音放大等問題與極點完美匹配的強烈假設。此篇論文的第二部分分析與討論了原有架構之失真的現象，並藉由後極濾波器作修正；同時探討了回音路徑之極點在極點偏差效應下對於共極點近端等化架構的影響，並比較了共極點近/遠端等化架構。此外，電腦模擬將會驗證上述演算法知可用性。

In stereophonic acoustic echo canceller (SAEC), the high cross-correlation between input signals from the far-end leads to slow convergence of the adaptive filter. In this thesis, two kinds of methods are studied to improve convergent rate: decorrelation-based algorithms and the common acoustical poles (CAP) structures. In the first part, two types of algorithms are proposed by integration of previous decorrelation algorithms to improve the convergence rate. In the second part, common-poles structures are studied to reduce the lengths of adaptive filters. A post-filter is used equalize the distorted near-end speech. The pole-mismatch effect is also studied. Computer simulations will be performed to validate the above algorithms.