Musical Audio Denoising Assuming Symmetric α-Stable Noise

The representation of α-stable distributions as scale mixture of normals is exploited to model the noise in musical audio recordings. Markov Chain Monte Carlo inference is used to estimate the clean signal model and the α-stable noise model parameters in a sparse linear regression framework with structured priors. The musical audio recordings were processed both as a whole and in segments by using a sine-bell window for analysis and overlap-and-add reconstruction. Experiments on noisy Greek folk music excerpts demonstrate better denoising under the α-stable noise assumption than the Gaussian white noise one, when processing is performed in segments rather than in full recordings.