Signal Deflation and Paraunitary Constraints in Spatio-Temporal Fastica-Based Convolutive Blind Source Separation of Speech Mixtures

The FastICA algorithm of Hyvarinen and Oja is a popular procedure for blind source separation of non-convolutive signal mixtures. Recently, two different extensions of this procedure have been proposed for convolutive blind source separation of speech and other signal mixtures. In this paper, we describe the major differences and compare the performances of these approaches, illustrating how signal deflation or coefficient orthogonalization is employed to maintain uniqueness of the separated system outputs for both synthetic convolutive mixtures and speech mixtures as recorded in real-room environments. Our numerical evaluations indicate that (a) coefficient orthogonality through paraunitary constraints provide more robust estimation behavior than least-squares signal deflation with unit-norm constraints, and (b) all-pass constraints can be used to improve the perceptual quality of the extracted speech when only signal deflation is employed.