亂碼序列之遞迴式回音消除器

Abstract

在大部分回音消除的問題中，雙向語音的發生一直是對回音消除的一種困擾。跟一般的噪音比起來，近端語音除了有著跟噪音一樣的無法預測性外，同時更有著比噪音更高的功率。這使得一些利用量測殘存回音來作為濾波器調整的一些方法如 RLS，NMLS 失去回音消除的能力 在本論文中我們將介紹一種maximal-length correlation（MLC）的回音消除器來解決雙向語音的問題。這種架構主要是在遠端語音中摻入一小功率的亂碼序列，並利用此一序列估計房間的響應來達到回音消除的目的。我們將對這一架構的回音消除效果作分析，並推導出最理想的濾波器的級數。 同時我們會根據此一架構發展出一種新的遞迴式回音相除器來增進其回音消除的效果。為了確保此一架構在遞迴的過程中不會發散，我們推導出一個充分的收斂條件。除此之外我們還針對此一架構在雙向語音的環境作分析，已確保其仍有解決雙向語音的能力。對於此一架構我們提出了幾種不同的演算法，並分析其所需硬體要求。這樣使得我此一遞迴式的架構在實現上有更多的彈性

In the acoustic echo cancellation double talk is always a problem, for it is like a disturbing noise. When the near-end speech occurs, it makes the echo canceller fail to trace the room response especially for some error feedback adaptive filter like LMS and RLS. In this thesis, we will introduce a maximal-length correlation (MLC) algorithm to overcome the double-talk problem. Two correlation methods, sequential and batch MLC, will be proposed. We will derive a close form for the ERLE performance of the MLC echo cancellation algorithm. From that a trade off between the bit length (computational cost) and SMR (speech quality) can be seen. The ERLE performance also depends on the filter length, so we will derive an optimal filter length that maximizes ERLE . Based on the MLC algorithm we will develop recursive and iterative MLC algorithms to improve the performance of conventional MLC. A bound of the convergence will be derived to ensure the recursive and iterative MLC work. We will also analyze their ERLE performance in double-talk to show that they out perform conventional MLC in double talk. Several kinds of the recursive and iterative MLC will be proposed with different hardware requirements. That will make the application become more flexible.