Ultra-Low Complexity Residue Echo and Noise Suppression Based on Recurrent Neural Network

Deep learning residue echo suppression (RES) exhibits superior performance compared with traditional methods in recent years. However, a low-resource system or preemptive multi-tasking system requires low-complexity model that should be very computationally efficient to reduce race-condition issues which could cause system delay jitter and echo delay changes. In this paper we do an extensive study on low-complexity recurrent network models with different topologies, feed in with different combinations of the far-end signal, microphone signal, predicted linear echo and residue error signal. The proposed RES models can achieve comparable echo cancellation and noise reduction capabilities to the AEC Challenge 2022 baseline model at a complexity lower than 5% of the baseline model.