Experiments on a parametric nonlinear spectral warping for an HMM-based speech recognizer

This paper is concerned with the search for an optimal feature-set for a speech recognition system. A better acoustic feature analysis that suitably enhances the semantic information in a consistent fashion can reduce raw-score (no grammar) error rate significantly. A simple two-dimensional parameterized feature-set is proposed. The feature-set is compared against a standard mel-cepstrum, LPC-based feature-set in talker-independent, connected-alphadigit HMM-based recognizer. The results show that a particular combination of parameters yields a significantly lower error rate than the baseline mel-cepstrum LPC-based feature-set