Enhanced wiener post-processing based on partial projection back of the blind signal separation noise estimate

In this paper, we consider the human/machine hands-free speech interface where the user voice is picked at a distance with a microphone array. It is assumed that the user is close enough to the machine to be considered as a point source. The noise is a diffuse background noise, created by all the sources present in the environment. The proposed method aims at suppressing the diffuse background noise efficiently without distorting the speech estimate. This method is a modification of a method combining frequency domain blind signal separation (FD-BSS) and Wiener filter based postprocessing. Contrary to the conventional approach, we modify the estimate of the diffuse background noise given by FD-BSS before applying the Wiener filter based post-processing. We also have to build a modified observation using the FD-BSS in order to apply our modified post-processing. Simulation results show that the proposed approach can achieve a better speech enhancement, measured in term of word recognition in a speech recognition task, than the conventional Wiener filter based post-processing.