Residual Echo Suppression Using Deep Learning

Adaptive Echo Cancellation (AEC) is a set of techniques meant to decrease the unwanted echo in an audio signal recorded by a communication device. Typically, noise and nonlinear distortions remain after AEC. Therefore, Residual Echo Suppressor (RES) is applied to improve upon AEC.
The goal of this project is to harness a deep learning approach to improve the RES.
Classical approaches that implement RES are far from being ideal. In this project, we ask the question: Does the deep learning approach perform better than the classical approaches?
To answer this question, we set up a data acquisition simple system, comprised of a microphone and a speaker. We collected a new dataset and trained several neural network (NN) architectures on it.
Our results show the feasibility of such a deep-learning RES approach. In some aspects, our RES outperforms other techniques.
We believe that better modeling of the actual communication scenario and a room acoustic response will help to acquire more precise dataset for training NN and, therefore, will improve results substantially.

Residual Echo Suppression Using Deep Learning
Residual Echo Suppression Using Deep Learning
Residual Echo Suppression Using Deep Learning