On Non-stationary Hidden Markov Modeling of Speech Signals

Hidden Markov, or Markov source, modeling of spech signals has long been applied to speech recognition. In particular, models with mixtures of Gaussian autoregressive (AR) output density functions have proved successful fro several applications. The commonly used approach [1] for performing this modeling assumes that a set of normalized autocorrelation vectors from the source, which comprises the training sequence, is given.