Feature selection in Mandarin large vocabulary continuous speech recognition

In this paper, three kinds of important features in speech recognition are introduced, including, LPCC, MFCC and PLP in order to compare the performances of these features, we design detailed experiments about the feature selection in a Mandarin large vocabulary continuous speech recognition system. In this system the semi-syllable based CHMM is applied to establish speech models and the all-syllables recognition accuracy is adopted to represent the performance. By employing different kinds of features, comparing diverse dimensions and selecting the computation methods of energy features, we utilize 45-dimension MF-PLP features in a Mandarin speech recognition system for the first time and obtain the best all-syllables recognition accuracy in published documents.