Robust treatment of impulsive noise in speech and audio signals

Markov chain Monte Carlo methods are presented for treatment of localized, impulsive noise (outliers) in digitized waveforms, within a Bayesian hierarchical framework. Outliers in audio signals occur as 'clicks' and 'crackles' in degraded sound recordings and impulsive noise in communications channels. Sampling-based methods for detection and correction of such artefacts are presented, in which individual noise sources are modelled as Gaussian with unknown scale, allowing for robustness to heavy-tailed noise distributions. Results are presented for speech and audio signals obtained from digitized sound recordings.