| Acoustic echo cancellation (AEC) is an essential component of effective hands-free telephony. Conventional AEC systems employ linear adaptive filters; therefore transmission channel nonlinearities caused by nonlinear components (especially the vocoders in digital networks) can severely degrade performance.; This dissertation examines the performance of popular conventional AEC algorithms based on adaptive filtering theory in nonlinear channels. In order to study the degradation of the algorithms in nonlinear channels, properties of nonlinear devices such as vocoders and loudspeakers are investigated, and a local linearization model is developed for the analysis of the nonlinear devices. This local linearization model is justified by experiments.; A variable step-size adaptive cross-spectral algorithm is proposed so the acoustic echo can be suppressed even during double-talk (DT) periods. This is important since adaptation is frozen during DT periods in order to avoid divergence of the conventional adaptive filtering algorithm; therefore the power of the residual echo may become higher than that of the original echo in nonlinear channels. In addition, the proposed algorithm does not need a DT detector, which is still part of AEC.; To compensate the echo attenuation loss of AEC algorithms caused by channel non-linearities, post-filtering techniques are exploited. Combined with a linear adaptive filter, post-filters based on various approaches, namely: Wiener-type post-filter, spectral subtraction, subspace method and pitch extraction, are proposed to further attenuate the echo. Experimental results show that the combined AEC system can suppress the acoustic echo to a satisfactory level in the nonlinear channel.; Subband adaptive filtering is also studied to reduce the computational complexity of the AEC system so that it can be implemented in real-time. To this end, an improved simple design of DFT filter banks is proposed. Furthermore, a post-filter is integrated with an adaptive filter in the subband to significantly suppresses the acoustic echo in the presence of channel nonlinearities. This approach also significantly reduces the computational burden.; Finally, a psychoacoustic approach based on the masking of the human ear is exploited in order to mitigate the artifacts resulting from the abovementioned post-filters. Testing indicates that the proposed method significantly reduces the distortion of near-end speech when DT occurs. This makes any audible residual echo sound more natural since it has less musical noise. |
