Robust noise modelling with application to audio restoration

New methods are presented for robust modelling of noise sources drawn from heavy-tailed or impulsive distributions, such as are commonly encountered in communications systems and corrupted audio signals. The methods are formulated for linear signal models within a Bayesian framework (although likelihood-based results are easily obtained as a subset of the Bayesian methods). Solutions are generated using powerful iterative techniques. Investigations are carried out for an audio interpolation framework in which certain samples are corrupted with additive impulsive noise.