Situated state hidden Markov models

The authors introduce a probabilistic model called a situated state hidden Markov model (SSHMM), in which states are situated (i.e., assigned positions) and assumed to correspond to regions of an underlying continuous state space. Transition probabilities among states are induced by the assigned state positions in such a way that transitions occur more frequently between nearby states. The model is formally defined, and a maximum likelihood estimation procedure is described. Experiments on synthetic data demonstrate the SSHMMs can learn the structure of an underlying continuous state space even when observed through high-dimensional discontinuous functions. Experiments using SSHMMs for speaker-independent phonetic classification are also reported.<>