Background

For a healthy person, it is an incredible feat to tune in to a single conversation in a crowded room while being completely oblivious to background noise and other conversations that happen simultaneously. This ability for a person to solve the “Cocktail Party Problem” is remarkable and very natural as a human. However, a significant percentage of the world’s population suffers from disabling hearing loss. For the affected people the ability to perceive speech especially in the presence of background noise is significantly impaired. Current hearing aids offer only limited relief as they cannot distinguish between speech and background noise and will simply opt to amplify both signals. This does not greatly help with the intelligibility. Hence, the focus should be on designing an “intelligent” system that can distinguish between the conversation and disruptive background noise in order to provide an improvement to hearing aids. It is not inconceivable that such systems would also be of great help to the people with a normal and healthy hearing considering the fact that listening to someone in a loud environment is still detrimental to the hearing. A system that can remove loud background noise but allow for unhindered conversation might, therefore, help prevent such cases of hearing loss.

This project contributes to these goals by comparing different implementations of supervised deep learning systems that can learn to enhance speech from background noise in monaural (single microphone) recordings.

Illustration: Cocktail Party Effect