Modeling Acoustic Transitions in Speech by Modified Hidden Markov Models with State Duration and State Duration-Dependent Observation Probabilities