Continuous speech recognition using syllabic segmentation and demisyllable hidden Markov models