ORIGINAL

Abstract

The task of speech recognition is challenging, as its applications rarely occur in ideal scenarios or environments. Practical implementations demand robust algorithms to address the complexities associated with real-life environments. One of these areas is acoustic echo cancellation (AEC), which attempts to mitigate the echo introduced by the acoustic coupling between a speaker system and a microphone within a full-duplex telecommunication system. We propose that a fully connected neural network is capable of modelling the complex nonlinear nature of acoustic echo.

This thesis is divided into three main components: a literature review of the state of the current field, a proof of concept model for a neural network based approach, and an implementation and comparative analysis to conventional linear adaptive filters. We find that neural network models are capable of outperforming conventional approaches to AEC, and our results open up opportunities for further investigation on the robustness of such models in real world applications.