Echo Cancelation and Noise Suppression by Training a Dual-Stream Recurrent Network with a Mixture of Training Targets

Nonlinear echo in presence of background noise can degrade the performance of digital signal processing algorithms. Deep neural networks with their ability to model complex nonlinear functions can potentially address this issue. In this paper, a deep and causal neural network based on dual streaming of the near-end microphone and far-end speech signals is employed to leverage the real-time nonlinear echo cancellation and noise suppression. The extracted features of two streams are coupled into a shared neural network for joint echo and noise cancellation. The training target is a mixture of spectral mapping and masking-based targets which are gated through a feedforward neural network. The model is evaluated in terms of both signal-level and perception-level metrics for different scenarios with a range of SI-SDR as low as −25 dB. Furthermore, the effect of mixing of training targets is assessed by evaluating different models.