Machine Learning Techniques for Speech Enhancement in Audio Conferencing Systems

The core value of Nureva is to provide reliable and easy-to-use audio-conferencing products that offer a good user experience and maximize productivity. A very important factor of that is to output clean audio which oftentimes in its raw form is contaminated with environmental noise, reverberation, and echo. These problems have been researched using traditional signal processing methods for a long time. While these can be quite effective in many situations, they do have limitations with non-stationary noises that vary with time. The success of Machine Learning (ML) in various fields also unlocks new possibilities in noise suppression and speech enhancement in that manual feature engineering is reduced or eliminated. That allows ML models to generalize to situations it has not encountered in training. This research aims to investigate the merits of various ML techniques and how they can be applied with Nureva products to ensure better audio performance.

Faculty Supervisor:

Nick Koudas

Student:

Partner:

Nureva Inc.

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

University of Toronto

Program: