An adaptive MEL-LPC analysis for speech recognition

This paper describes a new speech analysis method, an adaptive Mel-LPC (AMLPC) analysis method, using human auditory characteristics. The Mel-LPC analysis method that we have proposed is an efficient time domain technique to estimate the warped predictors from input speech directly. However, the frequency resolution of spectrum obtained by Mel-LPC analysis is constant regardless of the characteristics of input speech at each analysis frame. In the AMLPC analysis, it is probable to estimate the spectrum coefficients with optimal frequency resolution according to the characteristics of the phoneme at each analysis frame, because the spectral slope or the formant is different according to phoneme (vowels, fricatives and so on). The recognition performance of melcepstrum parameters obtained by the AMLPC analysis was compared with those of mel-cepstrum parameters obtained by the conventional LPC analysis and the Mel-LPC analysis through gender-dependent phoneme and word recognition. The results show that the proposed method leads to a significant improvement in recognition accuracy over conventional LPC analysis, and a slightly improvement of error rate about 10% over the Mel-LPC analysis.