Continuously variable duration hidden Markov models for automatic speech recognition