Bayes meets Bach: applications of Bayesian statistics to audio restoration

Abstract

Memoryless nonlinear distortion can be present in audio signals, from recording to reproduction: bad quality or amateurishly operated equipments, physically degraded media and low quality reproducing devices are some examples where nonlinearities can naturally appear. Another quite common defect in old recordings are the long pulses, caused in general by the reproduction of disks with deep scratches or severely degraded magnetic tapes. Such defects are characterized by an initial discontinuity in the waveform, followed by a low-frequency transient of long duration. In both cases audible artifacts can be created, causing an unpleasant experience to the listener. It is then important to develop techniques to mitigate such defects, having at hand only the degraded signal, in a way to recover the original signal. In this thesis, techniques to deal with both problems are presented: the restoration of nonlinearly degraded recordings is tackled in a Bayesian context, considering both autoregressive models and sparsity in the DCT domain for the original signal, as well as through a deterministic solution also based on sparsity; for the suppression of long pulses, a parametric approach is revisited with the addition of an efficient initialization procedure, and a nonparametric modeling via Gaussian process is also presented.