A framework for fast segment model by avoidance of redundant computation on segment

The segment model (SM) is a family of methods using segmental distribution rather than frame-based features (e.g. HMM) to represent the underlying characteristics of the observation sequence. It has been proved to be more precise than that of HMM. However, the high complexity prevents these models' use in practical systems. In this paper we present a framework to reduce the computational complexity of the segment model by fixing the number of the basic unit in the segment to share the intermediate computation results. Our work is twofold. First, we compared the complexity of SM with HMM and proposed a fast SM framework based on the comparison. Second we use two examples to illustrate this framework. The fast SM have better performance than the system based on HMM, and at the mean time, we successfully keep the computational complexity of SM at the same level as HMM.