A semi-blind approach to the separation of real world speech mixtures

The possibility of introducing a-priori information into multichannel blind deconvolution algorithms is investigated. The maximum likelihood (ML) approach allows one to introduce an important feature of the voice, namely the pitch, naturally into the 'blind' model, removing the nonlinearity and showing the advantages of productive contaminations by such related research fields as computer-aided sound analysis (CASA) and Bayesian theory.