A Bayesian Approach to Time-frequency Based Blind Source Separation

In this paper we propose a bayesian approach for time-frequency (t-f) based source separation. We propose a Gibbs sampler, a standard Markov Chain Monte Carlo (MCMC) simulation method, to sample from the mixing matrix, the source t-f coefficients and the input noise variance, under two models for the sources. In the first one the t-f coefficients of the sources are assumed i.i.d, while a frequency dependent modeling of the coefficients is proposed in the second one, which provides improved interference and noise rejection. Audio results are presented over several time resolutions of the t-f transform.