BAYESIAN SEPARATION OF DISCRETE SOURCES VIA GIBBS SAMPLING

Stephane Senecal and Pierre-Olivier Amblard
e-mail: Stephane.Senecal@inpg.fr, Bidou.Amblard@inpg.fr

Source separation, one of the most recent domain of sig- nal processing, consists in recovering signals mixed by an unknown transmission channel. Many approaches manage the separation using di erent techniques such as likelihood and information theory ([1],[6]) or high order statistics ([7]). This paper proposes a Bayesian approach to the problem of an instantaneous linear mixing, considering the sources are discrete. The powerful Gibbs sampling algorithm enables to recover binary sources, but also the mixing coe∆cients and noise levels with e∆ciency.