Dynamic Minimum Subband Spectral Subtraction and Its Application in Robust Speech Recognition

A nonlinear feature process algorithm called dynamic minimum subband spectral subtraction (DMSS) is described, inspired by amplitude spectra properties of noisy speech. This method does not require noise estimation and it is effective in dealing with both stationary and non-stationary noise. Its application for minimizing mismatch between clean and noisy speech features is also present. Experimental results show the proposed method can effectively improve the robustness of automatic speech recognition (ASR) and when combined with peak isolation method properly, it can improve the recognition performance greatly