Spectral feature-based nonlinear residual echo suppression

We propose a method for nonlinear residual echo suppression that consists of extracting spectral features from the far-end signal, and using an artificial neural network to model the residual echo magnitude spectrum from these features. We compare the modeling accuracy achieved by realizations with different features and network topologies, evaluating the mean squared error of the estimated residual echo magnitude spectrum. We also present a low complexity real-time implementation combining an offline-trained network with online adaptation, and investigate its performance in terms of echo suppression and speech distortion for real mobile phone recordings.