一种快速回声消除算法及DSP实现

A fast converging algorithm for network echo cancellation and its DSP implementation

在回声消除系统中,人们希望回声消除器不仅要收敛速度快而且要鲁棒性好。PNLMS和PNLMS++算法具有收敛速度快的优势,然而在双端发声情况下,它们就会产生大的失调,甚至不能保证滤波器收敛。为了克服上述问题,根据正交原理,我们提出了OP-PNLMS算法,该算法不仅具有PNLMS和PNLMS++算法的快速收敛性能,而且受双端发声和嘈杂背景噪声影响极小。即使是在单端发声时,该算法的性能也优于NLMS以及PNLMS和PNLMS++算法。此外,讨论了该算法在定点DSP上的实现方案。

In the applications of network communication, echo cancellers with faster convergence and more robust algorithms are highly expected. PNLMS （Proportionate Normalized Least Mean Square） and PNL-MS＋＋, which are recently proposed, are the algorithms to achieve more rapid convergence. However, PN-LMS and PNLMS＋＋ suffer from the same problem, such as ＂double-talk＂ disturbance, as most traditional algorithms do. To tackle this problem, this paper presents a novel algorithm named OP- PNLMS （Orthogo-nality Principle-Proportionate normalized LMS）. This algorithm not only converges as fast as PNLMS＋＋, but also shows more robustness than PNLMS and PNLMS＋＋ during double-talk or noisy background. Furthermore, OP-PNLMS outperforms PNLMS, PNLMS＋＋, and traditional NLMS even during single-talk. Several considerations of fixed-point DSP implementation of OP-PNLMS are also discussed in the paper.