A Unified Parameterization Scheme for Noisy Speech Recognition

Abstract

LP-based and mel-cepstrum coefficients are by far the most prevalent parameterization techniques in speech recognition. The conventional LP technique is known to be very sensitive to the presence of additive noise and there are few comparative studies about its relative robustness to noise with respect to mel-cepstrum. In this paper, a new unified parameterization scheme is proposed that combines both LP and mel-scaled filter bank analysis and yields new hybrid methods. Also it has been considered the frequency filtering of log filter-bank energies proposed recently by the authors as an alternative to cepstrum representations. From the comparison between conventional and new techniques, excellent results are presented in CDHMM clean and noisy digit recognition.