Denoising on adapted wavelet packets domain for robust speech recognition

Abstract

In a real environment, additive noise will corrupt input speech for speech recognition. In this paper, the authors propose a noise suppression method on the wavelet packet domain as a front-end pre-processor for robust speech recognition. They focus on the enhancement of the formant characteristic to input speech. Suppose, one has observations y/sub i/=f(t/sub i/)+ /spl sigma//spl middot/z/sub i/, i=0, 1, ..., n-1, where f(t/sub i/) is the speech signal and z/sub i/ is i.i.d. white Gaussian noise (AWGN). And assume that one has an available library L of orthogonal bases, such as wavelet packet bases. Using these assumptions, the authors enhance the formant characteristic as well as SNR by adjusting each node variance from adapted wavelet packet transform (AWPT) tree. Experimental result shows an enhancement of SNR from 3.58 dB to 8.66 dB. Also, phoneme recognition performance is improved more than 6%. It confirms the robustness of proposed noise suppression method against additive white Gaussian noise.