New architecture combining blind signal extraction and modified spectral subtraction for suppression of background noise

IWAENC2008: the 11th International Workshop on Acoustic Echo and Noise Control, September 14-17, 2008, Seattle, Washington USA.In this paper, we propose a new architecture for enhancing the speech in a hands-free human/machine communication scenario. The proposed architecture uses frequency domain blind source extraction (FD-BSE) for noise estimation and beamforming in the speech direction. Soft masks function of the SNR are computed from the FD-BSE outputs and use to enhance the performance of the spectral subtraction applied channel wise to suppress the noise. Simulation results show that the proposed architecture can achieve a comparable SNR as conventional spectral subtraction with less distortion of the speech