Bayesian blind separation of audio mixtures with structured priors

In this paper we describe a Bayesian approach for separation of linear instantaneous mixtures of audio sources. Our method exploits the sparsity of the source expansion coefficients on a time-frequency basis, chosen here to be a MDCT. Conditionally upon an indicator variable which is 0 or 1, one source coefficient is either set to zero or given a Student t prior. Structured priors can be considered for the indicator variables, such as horizontal structures in the time-frequency plane, in order to model temporal persistency. A Gibbs sampler (a standard Markov chain Monte Carlo technique) is used to sample from the posterior distribution of the indicator variables, the source coefficients (corresponding to nonzero indicator variables), the hyperparameters of the Student t priors, the mixing matrix and the variance of the noise. We give results for separation of a musical stereo mixture of 3 sources.