利用快速相等 NLMS 作分頻回音消除

Abstract

摘要: 由於適應性回音消除器的係數通常高達幾千個, 因此造成收 斂速度慢和運算複雜度高的問題. 本篇論文提出一個快速相等 NLMS 的演 算法, 它和 NLMS 有相似的收斂性但其運算複雜度較低. 對語音信號輸 入, 提出另一個演算法 DFENLMS 來改善 FENLMS 的收斂性.分頻適應性濾 波可以降低運算複雜度並增加收斂速度. 然而分頻濾波會有不可避免的重 疊問題, 因而造成分頻回音消除的殘餘回音信號上會有一個尖端成份. 我 們用陷波濾波器來消除這個尖端成份並且設計不同的陷波濾波器. 我們分 析了不同的陷波濾波器對系統表現和近端語音的影響. 最後我們提供能減 少乘法器的多項法來實現濾波器組和陷波濾波器. Abstract: Thousands of tap weights of an adaptive FIR echo canceller usually results in slow convergence speed and high computational complexity. In thisthesis, a fast equivalent NLMS (FENLMS) algorithm is proposed which has lower computational complexity and similar convergence characteristic as compared tothe NLMS algorithm. For speech signal, another decorrelation FENLMS (DFENLMS) algorithm is proposed to improve the performance of the FENLMS algorithm.Adaptive filtering in subband can reduce the computational complexity and increases the convergence speed. Howerer, aliasing is an unavoidable problemin subband filtering, and a peak component will appear in the residual echo signal. We use notch filters to cancel this peak component and different notch filters are designed. Analysis of differnet notch filters on system performanceand andnear-end speech are given. Finally, the polyphase implementations of filterbanksand notch filters are given to reduce the number of multipliers.