During my PhD I worked on the simultaneous removal of impulse and background noise from audio signals. Impulse noise is most often heard as the clicks and pops in old vinyl recordings, whereas background noise is most commonly a fairly uniform ‘hiss’ on the background of a track. Our method used a Gabor wavelet decomposition to remove the background noise, whilst using a form of sparse Bayesain inference to determine whether a particular audio sample was part of an impulse. The method proved to be successful in the removal of the target noise. Since the method is computationally intensive and works in a batch form (meaning the full track, or at least several seconds are required at once), in its present form it is best suited to audio restoration rather than real-time sequential processing, though this could be addressed in future work.
This work led to the paper Joint Bayesian removal of impulse and background noise, which I presented at the ICASSP signal processing conference in Prague in May 2011.