利用相關函數偵測之可變步伐回音消除器研究

Abstract

在回音消除的領域中，熟悉的NLMS演算法被廣泛的使用。然而，使用固定收斂步伐的傳統NLMS演算法，具有收斂速度與位錯 (Misa-lignment) 之間的權衡問題，因此許多的可變步伐演算法被提出。此外，當雙邊通話發生時，會造成適應性濾波器係數發散，因此也有許多的雙邊通話偵測器被提出。 在此篇論文中，我們首先介紹Huang所提出的可變步伐演算法，並進行暫態步伐與位錯的分析，最後再探討房間響應改變偵測器。接著，我們介紹與提出各種不同的基於相關函數之雙邊通話偵測器與近端語音能量估測器，並進行理論的分析與比較。另外，我們也提出兩種方法可以使雙邊通話偵測器正確的辨別雙邊通話與房間響應改變。第一種是採用雙路徑的架構，然而其需要二個濾波器；第二種則是採用有限狀態機的架構，其只需要一個濾波器。最後，電腦模擬將驗證理論分析與提出的方法。

In the field of acoustic echo cancellation, the well-known NLMS algorithm is widely used. However, the conventional NLMS algorithm with the fixed step-size has trade-off between convergence rate and misalignment. Thus, many variable step-size (VSS) algorithms have been proposed. Furthermore, the coefficients of adaptive filter diverge when double talk occurs. To solve this problem, many double talk detectors (DTD) have also been proposed. In this thesis, we first introduce Huang’s VSS algorithm and perform transient analysis of step-size and misalignment. Then, we discuss the echo path change detectors. After that, we introduce and propose different correlation-based DTDs and near-end speech energy estimators, and also perform the theoretical analysis and comparisons. Moreover, we propose two kinds of methods to distinguish DT and EPC correctly for DTD. The first one uses the two path structure, however it requires two filters; the second one uses the finite state machine and only one filter is required. Finally, the simulation results will be presented to verify our discussions.