Spatial coherence exploitation which yields non-stationary noise reduction in subband domain

This paper deals with the enhancement of speech corrupted by real additive noises in a car when two observations are available. As far as we know, no enhancement system was capable of improving both the quality and the intelligibility of the noisy signals. We propose an enhancement method using thresholding, segmentation and filtering in subband domain. The main idea is to expand in subband signals the two observation speech signal and to exploit the spatial coherence of sound from the subband expansion. Noise reduction is conducted using two methods according to the degree of correlation of the subbands of the two observations. The proposed noise reduction approach is applied to realistic situations like speech signals, which are corrupted by non-stationary noises of diverse origins.