Acoustic Echo Detection using Deep Learning

Abstract

In acoustic echo cancellation, adaptation of echo estimation filter relies on the detection of the state of acoustic echo canceller (AEC). Conventional method of detection of echo only state has high probability of false detection or miss detection, which results into divergence or slow convergence of the normalized least mean square (NLMS) algorithm. In this work, we focus on enhancing the accuracy of the echo only state detection using deep learning algorithms instead of conventional detector. We prepare data set from the speaker and microphone signals to train the deep learning algorithms. We use Alex Net, Deep convolution neural network (DCNN), Recurrent neural network (RNN) and K-nearest neighbor(KNN) for the echo detection. We prepared two training data sets each containing 2000 echo samples and 2000 samples without echo. Two testing data sets contains 60 echo samples and 40 no-echo samples. The aforementioned algorithms are trained on data sets, then tested on testing data sets achieves promising results. This trained echo detector detects echo only state and helps achieve better convergence of the NLMS adaptive filter.