雙聲道迴音消除之研究

Abstract

對於多人的視訊會議，採用多聲道可使聲音更傳真，利用每個人位置不同的特性，多聲道可使聽者更容易辨別誰是說話者。最簡單的多聲道為二個聲道，即所謂的立體聲。對於免持聽筒系統，迴音消除(AEC)是很重要的機制，否則講話者會聽到自己的迴音。立體聲之擴音迴音消除，和單聲道相比，衍生出更多的問題。主要的問題在於二個聲道彼此有很強的相關性，導致使用適應性濾波器來模擬二個聲道的房間響應時，並不能模擬得很像。本篇論文研究著重於單一聲道到雙聲道的退化效應，建立單聲道與雙聲道誤差向量的數學關係式，並分析雙聲道誤差向量放大的因子。我們也考慮遠端說話者改變，近端濾波器參數調整的行為，分析迴音量的影響。在適應性演算法方面，我們研究雙聲道LMS演算法的行為，並和單聲道LMS作比較，測試兩種改善雙聲道迴音消除的方法（非線性處理及全通時變濾波器），進而分析採用時變濾波器，誤差向量在LMS的二階統計特性的收斂行為。最後我們提供加入虛擬噪音（PN sequence）的方法以求出一個迴音路徑，同時聲音的位置特性並不會被改變。

Recently, spatial realism is desirable for vivid audio/video teleconference systems because listeners can use spatial information to help distinguish who is speaking. This need leads to consideration of stereophonic sound. Based on the Wiener-Hopf equation in stereophonic AEC, we build the mismatch vector relation between single channel and dual channel and analyze the degradation effect in the case of dual-channel with respect to that of a single channel. We also analyze poor misalignment in stereophonic AEC as well as increased echo power when the far-end room impulse responses change. Compared with conventional mono channel case, we also investigate the convergence of the stereophonic LMS algorithm. We also consider two decorrelated approaches, nonlinear filters and time-varying all-pass filters, and perform convergence analysis of the misalignment power using time-varying filters in stereo LMS algorithm. Finally, we propose to add a PN sequence in one channel so that we can identify echo paths without changing spatial information.