Generation and Minimization of Word Graphs in Continuous Speech Recognition

The ability of an automatic speech recognition system (ASR) to output multiple recognition hypotheses is becoming increasingly important. For example in dialogue systems and machine translation systems where the ASR must interact with modules that handle discourse, semantics and pragmatics. Passing multiple recognition hypotheses to the higher levels of the system is a way of increasing the coupling without losing the modularity, which is essential for the development of large complex systems. Another application is hypothesis re-scoring, where the multiple hypotheses are an internal, intermediate representation generated by a fast initial search. They are then used to limit the search-space in a second, more accurate search.