Maximum likelihood successive state splitting

摘要: Modeling contextual variations of phones is widely accepted as an important aspect a continuous speech recognition system, and much research has been devoted to finding robust models context for HMM systems. In particular, decision tree clustering used tie output distributions across pre-defined states, successive state splitting (SSS) define parsimonious topologies. We describe new design algorithm, called maximum likelihood (ML-SSS), that combines advantages both these approaches. Specifically, topology designed using greedy search the best temporal splits constrained EM algorithm. Japanese phone experiments, ML-SSS shows performance gains training cost reduction over SSS under several conditions.