Background noise classification and denoising of audio signals utilizing empirical wavelet transform and deep learning

Abstract

The importance of sound and speech cannot in human life cannot be overstated. Ambient sound informs us about our surroundings and warn us of potential dangers, such as a car approaching from behind. One of the most common modes of communication is speech. However, if it is contaminated with background noise, it may result in data loss. Recent advances in artificial intelligence enable machines to recognize and classify sound patterns, as well as remove complex background noise from contaminated speech.

This thesis investigates a method for improving existing background noise classification and denoising solutions. This is accomplished through signal decomposition with the Empirical Wavelet Transform and subsequent processing with the Convolutional Neural Network. Improvements of up to 18% have been observed.