Subband acoustic echo cancellation

Abstract

The main theme of this thesis is the control of acoustic echoes for modem voice communication systems by means of echo cancellation. Two important issues in acoustic echo cancellation, namely the efficient adaptation of the echo cancellation filter and the reliable adaptation of the echo cancellation filter in double talk environment, are investigated. The delayless subband adaptive filter architecture is studied. Efficient implementation of the analysis filter bank and the time domain filtering are derived. The transforming of the subband filter weights to a fullhand counterpart is examined. It is shown that the weight transform is a synthesis filtering procedure. Two new weight transform schemes that deliver substantial performance improvements are proposed. The open-loop optimal subband filter impulse responses are shown to be non-causal and several anti-causal laps in the subband filters are required to model this non-causality. Because of the inevitable double talk detection errors, adaptive filtering algorithms with built-in double talk robustness measures are needed for the reliable operation of the echo canceller. The basic idea of robust adaptive filtering is examined. A comparison of different existing time domain robust adaptive filtering algorithms demonstrates that excellent trade-off between the convergence and the tracking properties and the double talk robustness of the adaptive filtering algorithm can he achieved by using Huher’s method for both the update of the echo cancellation filter and the estimation of scale. A delayless closed-loop robust sub- hand adaptive filter is proposed.By independently adapting the scale estimates and normalizing the adaptation in each subband, significant improvement in terms of the convergence and tracking speed over the time domain robust NLMS algorithm can be obtained without sacrificing the double talk robustness. Moreover, it is demonstrated that by using different thresholds in the update of the echo cancellation filter and the scales, the robust algorithms converge and track echo path variation as fast as their non-robust counter part while still maintaining a sufficiently low sensitivity to double talk detection errors. The application of two path adaptive filters to acoustic echo cancellation is examined. An analysis of the original two path adaptive filtering algorithm shows that it suffers from two kinds of performance degradation due to the divergence of the background filter during double talk, namely the slow tracking of echo path variation and the false filter coefficient copying after double talk. A robust two path adaptive filter is proposed to mitigate these problems.