Efficient blind speech signal separation combining independent component analysis and beamforming

Abstract

Blind Source Separation (BSS) algorithms can increasingly separate speech signals utilizing time and frequency domain information and beamforming (BF ) algorithms using spatial information from different point of views. BSS exploits a strong statistical condition including independence between source signals, while the popular BF approach utilizes the spatial information about the mixing system and/or source signals. In beamforming stage, the Directions of Arrival (DOA) of sources of interest are estimated blindly and then beamformers are constructed to extract signals from these directions. In the BSS stage, frequency domain convolutive algorithm is utilized to further reduce the interference in the given direction and improve the separation performance. Compared with existing systems, the proposed approach significantly reduces the computational complexity while maintaining comparable separation performance.