Application of semi-Markov models to automatic speech recognition

Automatic speech recognition using a statistical pattern recognition technique called hidden Markov modeling is introduced. With this technique each word in a vocabulary is represented as a probabilistic function of a Markov process, or hidden Markov model. A hidden Markov model consists of an underlying Markov chain, plus a set of multivariate probability density functions, one for each node of the Markov chain, called the states of the model. Intuitively the states describe variations in the pronunciation of different regions of a word while the underlying Markov chain describes the temporal structure of that word. Thus in hidden Markov modeling a speech signal is treated as the output of a sequence of stationary stochastic processes.