High-order hidden Markov modelling
Recently we introduced algorithms for the efficient processing of high-order hidden Markov models (HMMs) of fixed order. In this paper these techniques are generalized to also include mixed-order HMMs. This then allows first-order equivalents to be found for all HMMs, thereby providing a unifying base for reasoning about their properties. Specifically we provide formulations that separates the duration modelling and context modelling capabilities of high-order HMMs.