Using LVQ to enhance semi-continuous hidden Markov models for phonemes