Blind Source Separation Using Graphical Models

Abstract

We summarized our approaches for separating voices from mixed recordings. In the single channel case, a priori learned basis functions are used to model the temporal structure of the speech signals. A maximum likelihood approach is used to separate a voice from jazz music given only one mixed channel. In case of two microphones, the problem of separating two voices recorded by two microphones has been tackled. The mixing coefficients, time delays and reverberation coefficients are estimated using the maximum likelihood or infomax approach. The two approaches can be combined in a graphical model since both methods can be represented as data generative models where learning involves the representation of signals via the basis functions and inference involves the estimation of sources. The inference part in case of the single channel is nonlinear and linear in the two channel case.