Blind deconvolution of music signals using Higher Order Statistics

A method for the blind deconvolution of music recordings using Higher Order Statistics (HOS) is presented. Music signals can be modelled as sinusoids with noise. The noise part is assumed to have a nonGaussian statistics with a nonzero skewness. I show that when the 3rd-order statistics of a reverberated music signal is calculated, the effect of the deterministics part is cancelled and only noise convolved with the room impulse response (RIR) is observed. Therefore, using system identification methods based on 3rd-order statistics, RIR can be obtained and used to remove the reverberation. Simulations performed with real RIR and music signals confirm the method and validity of the ideas.