Separation of excitation source and vocal tract transfer function via an MVDR analysis of speech

The usage of MVDR (minimum variance distortionless response) spectrum estimation for separating excitation source from vocal tract transfer function in speech analysis is discussed. It is shown how an MVDR all-pole model and its inverse filter for source deconvolution can easily be derived by applying spectral factorization. We outline the theory of the Bauer method for spectral factorization, presenting examples of the application to real data recorded by distant microphones and giving insights on this technique.