Recognition of noisy speech using normalized moments

Spectral subband centroid, which is essentially the first-order normalized moment, has been proposed for speech recognition and its robustness to additive noise has been demonstrated before. In this paper, we extend this concept to the use of normalized spectral subband moments (NSSM) for robust speech recognition. We show that normalized moments, if properly selected, yield comparable recognition performance as the cepstral coefficients in clean speech, while deliver a better performance than the cepstra in noisy environments. We also propose a procedure to construct the dynamic moments that essentially embodies the transitional spectral information. We discuss some properties of the proposed dynamic features.