Derivation of Robust Mel-Frequency Cepstral Coefficients Using a Weighting Discrete Cosine Transform

Abstract

Mel-frequency cepstral coefficient (MFCC) is one of the most popular speech features used in an automatic speech recognition system. In order to improve its discrimination capability and robustness in various environments, a weighting discrete cosine transform (WDCT) is proposed in this paper and incorporated into the derivation of the conventional mel-frequency cepstral coefficients. The weighting function used in the discrete cosine transform can be easily calculated from the log-spectral amplitudes of each speech frame and by which we can adequately explore the relative reliabilities among different critical band filters. Experimental results for recognition of continuous telephone speech indicate that the syllable recognition rates of the WDCT-MFCC are 3.91%and 2.16% higher than those of the conventional MFCC in the cases of with and without compensation of channel distortions, respectively. Those results verify the robustness and effectiveness of the proposed WDCT-MFCC. Moreover, comparisons of F-ratio measures between the conventional MFCC and WDCT-MFCC also conclude that the WDCT-MFCC has superior discrimination capability in modeling a speech recognizer.