On Efficient Viterbi Decoding For Hidden Semi-Markov Models

We present algorithms for improved Viterbi decoding for the case of hidden semi-Markov models. By carefully constructing directed acyclic graphs, we pose the decoding problem as that of finding the longest path between specific pairs of nodes. We consider fully connected models as well as restrictive topologies and state duration conditions, and show that performance improves by a significant factor in all cases. Detailed algorithms as well as theoretical results related to their run times are provided.