Techniques to achieve fast lexical access

Three techniques are proposed to speed up the search of the lexicon in large-vocabulary continuous speech recognition. Using a language model heuristic allows the search to be nearly independent of the size of the language model and also brings the language model capability early in searching the lexicon. Local envelope pruning is implemented to reduce the number of candidates of lexical theories during the lexical access. A depth-first search is used to eliminate the overhead of the stack management in a stack search. These techniques have been shown to be effective on the Wall Street Journal speaker-independent task, and the 30,000-word system runs currently in real-time on a Sun workstation.