Speech recognition of mandarin syllables using both linear predict coding cepstra and Mel frequency cepstra

This paper is to compare two most common features representing a speech word for speech recognition on the basis of accuracy, computation time, complexity and cost. The two features to represent a speech word are the linear predict coding cepstra (LPCC) and the Mel-frequency cepstrum coefficient (MFCC). The MFCC was shown to be more accurate than the LPCC in speech recognition using the dynamic time warping method. In this paper, the LPCC gives a recognition rate about 10% higher than the MFCC using the Bayes decision rule for classification and needs much less computational time to be extracted from speech signal waveform, i.e., the MFCC needs computational time 5.5 time as much as the LPCC does. The algorithm to compute a LPCC from a speech signal much simpler than a MFCC, which has many parameters to be adjusted to smooth the spectrum, performing a processing that is similar to be adjusted to smooth the spectrum, performing a processing that is similar to that executed by the human ear, but the LPCC is easily obtained by the least squares method using a set of recursive formula.