Coherent time modeling of Semi-Markov models with application to real-time audio-to-score alignment

This paper proposes a novel insight to the problem of duration modeling for recognition setups where events are inferred from time-signals using a probabilistic framework. When a prior knowledge about the duration of events is available, Hidden Markov or Semi-Markov models allow the setting of individual duration distributions but give no clue about their choice. We propose two criteria of temporal coherency for such applications and prove they are fulfilled by statistical properties like infinite divisibility and log-concavity. We conclude by showing practical consequences of these properties in a real-time audio-to-score alignment experiment.