SNR-dependent non-uniform spectral compression for noisy speech recognition

Abstract

It is known that the perceived loudness of a tone signal by a human is spectrally masked by background noise. This masking effect causes not only a shift of just-audible sound pressure level of the tone, but also produces a masked loudness function having steeper slope than the unmasked one. This masking property of perceived loudness stimulates us to propose a new mel-scale-based feature extraction method with non-uniform spectral compression for speech recognition in noisy environments. In this method, the speech power spectrum is to undergo mel-scaled band-pass filtering, as in the standard MFCC front-end. However, the energies of the outputs of the filters are compressed by different root values defined by a compression function. This compression function is a function of the SNR in each filter band. Using this new scheme of SNR-dependent non-uniform spectral compression (SNSC) for mel-scaled filter-bank-based cepstral coefficients, substantial improvement is found for recognition in different noisy environments, as compared to the standard MFCC and features derived with cubic root spectral compression.