Acoustic Echo Cancellation System Using Adaptive Filtering Algorithm

Abstract

The field of acoustic echo cancellation (AEC) has been one of the key research area newlineaimed at improving the quality of acoustic signals [1,3,4]. The process of deep newlineresearch and thorough analysis of the acoustic signals has been aided by the newlineadvancement achieved in the recent days in the field of signal processing and newlineadvanced features development in the digital processing. The two areas pose an easier newlineway of extracting the possibilities of echo reduction in the waveforms in newlinecommunication. newlineThe advancement has brought forward several echo reduction techniques based on newlineadaptive filters that includes algorithms such as least mean square algorithm newlineabbreviated as LMS, Recursive least square (RLS) algorithm along with modified newlineform of LMS referred to as NLMS(Normalized least mean square) etc. Machine newlinelearning adaptive methods have become a popular research topic in speech newlineenhancement like echo cancellation. As a result of training these networks for certain newlinearchitecture they are expected to produce better performance than the conventional newlinefilters. In this paper, a two layered neural network (NN) architecture in combination newlinewith descent coordinate Iteration (DCI) modified RLS adaptive filter is proposed to newlineaddress both linear and nonlinear echo scenarios. The filter architecture proposed is newlinetrained in the frequency domain that follows the prediction of time-frequency mask newlinefor the target speech as per the error signal received. Results suggest that the proposed newlinemethod outperforms the existing traditional RLS method in terms of echo return newlinecorrelation coefficient enhancement and reduction in standard deviation along with newlinemean square error (MSE). Moreover, multiple analysis with changes in filter newlineparameter to suit more practical scenarios has also been deployed and results favored newlinethe proposed learning method. The comparative graphs for the two designed filter has newlinebeen depicted in the graph 5.8 and there has been enhancement seen in the parameters newlinealso. The standard deviation has been reduced to 0.2437 and the mean square e