ISSN: 2277-3754 ISO 9001:2008 Certified International Journal of Engineering and Innovative Technology (IJEIT) Volume 3, Issue 2, August 2013 246 Abstract—In this paper, an independent component analysis (ICA) acoustic echo cancellation (AEC) algorithm is introduced where a sliding discrete Fourier transform window is adopted such that there is only one AEC parameter to estimate (reduced computational load), as opposed to thousands of coefficients modeling the room response. Conventional adaptive filtering techniques such as the least mean square (LMS) algorithm often fail under double-talk condition (and excessive noise) due to a corrupted measure of the objective function (i.e. minimization of the error output). Recent study has shown that ICA allows continual adaptation of the AEC parameters, hence it is adopted here as the optimization method of our AEC parameter. Simulation results are used to illustrate the superiority of the proposed algorithm over the LMS methods. Index Terms—acoustic echo cancellation, blind deconvolution, double-talk detection, mutual information. I. INTRODUCTION The current acoustic echo cancellation (AEC) algorithms are based mostly on adaptive filtering techniques [1-3]. The loudspeaker signal is filtered by the loudspeaker-environment-microphone (LEM) impulse response to give the far-end signal , where is the convolution operator. The talker signal is filtered by the talker-environment-loudspeaker (TEM) impulse response resulting in the near-end signal . These two signals along with the observed room noise are captured by the microphone to give . In the absence of the near-end signal, an adaptive filter is used to model the LEM impulse response to give the far-end estimated signal . The estimated signal is then subtracted from the microphone signal resulting in the error signal , where is the residue signal. Assuming minimal room noise, the estimated response can converge to the desired LEM response and the echo can be effectively cancelled. The problem arises on observation of the near-end signal as now the filter will diverge resulting in poor echo cancellation. To circumvent the problem of filter divergence due to the near-end signal, double talk detection (DTD) algorithms such as cross-correlation methods [4, 5], the Geigel algorithm [6], and other variants are employed to detect the presence of the near end signal, in which case the adaptive process of modeling the LEM filter is frozen. Assuming minimal changes in the LEM enclosure, the far-end signal can still be effectively suppressed. Unfortunately, there are cases where the near-end signal is observed for long durations and since in most cases it is the source of this signal that controls the location of the microphone, changes in the LEM enclosure and hence the filter are far too frequent. This implies that the current estimate of the LEM filter in the adaptive process is inadequate to effectively reduce the residual signal and echo cancellation fails. In this paper, we propose the use of blind source separation (BSS) based on a blind deconvolution algorithm using a single frequency bin to address the problem a changing LEM filter in the presence of the near-end signal. By employing source separation as opposed to suppression of the far-end signal, there is no need for a DTD algorithm as the signals and can simply be separated, after which a mutual information check can be performed between each of the separated signals and loudspeaker signal to identify which is the echo (far-end signal) and the desired signal for transmission (near-end signal). The key point is that the filter modeling should not be frozen because of the near-end signal, and this is possible if BSS is employed. This paper is structured as follows: Section II introduces the proposed echo cancellation method. Simulation results for echo cancellation in the presence of the near-end signal for the proposed algorithm are in Section III. Discussions and summary remarks follow in Section IV. II. THE PROPOSED AEC ALGORITHM The proposed AEC algorithm is illustrated in Fig. 2. The loudspeaker signal is filtered by the LEM impulse response such that the far-end signal is , whereas the near-end signal is given by . Fig. 1. The loudspeaker and microphone signals are used as inputs the source separation algorithm of convolutive mixtures. The sources are identified using the correlation coefficient. DTD-free AEC via a Sliding DFT Window for ICA-based single Parameter Estimation E. S. Gower, T. Tsalaile, M. Kgwadi, S. Masupe