Denoising Audio using Autoencoders

Abstract

The problem of denoising has been an integral part in the landscape of audio processing for some time. Multiple approaches to this problem have been studied, such as Wiener filters, spectral noise gates, nonnegative matrix factorization, and deep neural networks. This thesis focuses on the latter approach, and aims to study how audio denoising can be achieved using autoencoders: a special type of neural network designed to learn an efficient data representation in an attempt to copy its input to its output. In this work, various autoencoder architectures are considered. The aim of this work is to compare the performance of different types of autoencoders in solving the denoising problem for audio signals, and use these findings to derive an autoencoder-based denoising method.