Removing noise from audio recording using Online Non-negative Matrix Factorization

Abstract

Dictionary learning has been shown to be an effective tool for signal processing. In this thesis, we look at a specific version of dictionary learning called Online Non-negative Matrix Factorization (ONMF) and apply it in the context of denoising musical recordings. We begin with a theoretical overview, highlighting the motivation for ONMF, such as its ability to train on a dataset while only ever requiring a small part of it to be loaded into memory. We then experimentally show that the denoising performance of ONMF depends strongly on the types of signals being processed and, to a lesser degree, on the correct choice of dictionary size for each signal.