Parallel and real-time implementation of an acoustic echo canceller using oversampled wavelet frame algorithms

Abstract

This thesis describes a novel echo cancellation system that eliminates nonstationary echoes with long acoustic delays in real-time. By combining subband adaptive filtering and active system identification based on fast wavelet transform and oversampled frames, the proposed architecture obviates the need for lengthy filters commonly found in conventional acoustic echo cancellers. Besides improving convergence due to subband decorrelation, wavelet decomposition offers a "divide--conquer" approach to meet the stringent filter length requirement. To cope with time-varying echoes, a channel probing engine is used to reset the adaptive filter array based on detection of returned pilot signals, using oversampled frame decompositionThe proposed architecture is then implemented on a parallel TMS32OC40 computer. Simulation results show that the hybrid architecture consistently outperforms both the conventional fullband and subband filters, even when the pilot signals are heavily corrupted by channel noise,