Hidden semi-Markov models (HSMMs).pdfVIP

Hidden semi-Markov models (HSMMs) Kevin P. Murphy /murphyk 20 November 2002 Abstract A semi-Markov HMM is like an HMM except each state can emit a sequence of observations. In this paper, we provide a unified review of inference and learning in a variety of different kinds of semi-Markov HMMs, including explicit-duration HMMs, segment HMMs, and multigrams. We also discuss some important implementation details, such as avoiding underflow, which are often glossed over. 1 Segment models 1.1 Representation A semi-Markov HMM (more properly called a hidden semi-Markov model, or HSMM) is like an HMM except each state can emit a sequence of observations. Let be the subsequence emitted by “generalized state” . The “generalized state” usually contains both the automaton state, , and the length (duration) of the segment, . We will define to be the subsequence . After emitting a segment, the next state is , where . Similarly, denote the previous state by . Let be all observations following , and be all observations preceeding , as in Figure 1. Each segment can be an arbitrary distribution. If ,