Acoustic echo and noise cancellation schemes using time and frequency domain adaptive techniques

Abstract

Acoustic echo occurs in real life environment when speech signal coming out from a loudspeaker is delayed, attenuated and reflected back to the source microphone. Most communication systems are prone to acoustic echo which can severely degrade the quality and intelligibility of the signals transferred through the communication channels. In conventional acoustic echo cancellation (AEC) methods gradient based adaptive filter algorithms, such as least mean squares (LMS) and normalized LMS are employed where an error function is minimized to obtain the optimal filter coefficients corresponding to the acoustic echo path. The main problem of these methods is the necessity of the dual channels, one for the reference signal and the other for the echo corrupted signal. However in many practical applications only one channel is available, such as a conference hall environment with single microphone and a loudspeaker. Due to the unavailability of separate reference signal in single channel scenario, the task of echo cancellation becomes extremely difficult and is attempted by a few researchers. In this thesis, first a single channel echo cancellation scheme is developed based on the gradient based LMS adaptive filter algorithm, where, unlike conventional dual channel schemes, a delayed version of an estimated echo cancelled signal is utilized as a reference signal. In the proposed formulation, the effect of flat delay, i.e. the time required to produce an echo, is incorporated with a view to reduce the number of unknown parameters of the acoustic echo path, which offers a faster convergence. Moreover, based on energy and cross-correlation coefficients of the reference and current frames, a multi-step stopping criteria is developed, which can efficiently control the updating procedure of the proposed LMS adaptive filter. Extensive experimentation is carried out on real life speech signals corrupted by echoes using the proposed single channel LMS algorithm with and without the multi-step update constraints. It is found that the performance of former one, the controlled LMS algorithm, is far better than that of the later one in terms of (a) the average echo return loss enhancement (ERLE) in dB and (b) the difference between input- and output-signal to distortion ratio (SDR) in dB. In real life applications, inclusion of noise with the speech signals is obvious in most of the cases, which makes the task of single channel echo cancellation even more difficult. In view of handling the challenging task of cancelling the echo in the presence of noise, a two step algorithm is developed where a spectral subtraction based noise reduction scheme is introduced after the single channel echo cancellation. It is shown that even under severe noisy conditions in different acoustic environments the proposed two-step single channel acoustic echo and noise cancellation (AENC) method can significantly reduce the effect of both echo and noise. As an alternate to the gradient based approaches, the problem of dual channel echo and/or noise cancellation can also be realized based on some optimization algorithm driven adaptive filters. However, undoubtedly the problem would be very difficult for the single channel scenario which is the case under consideration. Thus, the single channel AEC problem is formulated as an optimization problem introducing the particle swarm optimization (PSO) algorithm, which offers a quick convergence to the desired solution. For proper operation of the PSO algorithm, a frame by frame processing is required for which the overlap-add method is adopted. In order to estimate the unknown coefficients of the acoustic echo path, the PSO based algorithm is formulated both in the time and frequency domain separately and it is found that the frequency domain approach performs better in comparison to the time domain approach. The performance of the proposed PSO algorithms are also investigated for different controlling parameters, namely number of particles, maximum particle velocity etc. The PSO based algorithm is also extended for the complicated case of adaptive echo and noise cancellation. From detailed simulations it is found that the performance of the proposed PSO based AENC algorithm outperforms that of the proposed gradient based algorithm under different noisy conditions at various acoustic environments.