A new speech enhancement method based on LCMP combined with subspace theory

This article bring forward a new speech enhancement method based on local cosine bases matching pursuit (LCMP) combined with subspace theory. LCMP method is adaptive, through finding atoms matching the signal from an over-complete dictionary, this method could reconstruct the speech signal and remove the noise. However, for low SNR speech signal, LCMP method may appear to be mistaken issues estimating signal atoms as noise component atoms. The solution is bring in subspace theory. Firstly, isolating the noise subspace, and then in the signal subspace, making LCMP decomposition. Experimental results show that, this new method is fit for speech signals´ enhancement in various SNRs, and effectively suppress the music noise.