Incorporating Lr Parsing into Sphinx

This paper describes the integration of an LR natural language parser with the SPHINX speech recognition system. SPHINX is one of the most successful speech recognition systems in use today. Although it attains high word accuracy, SPHINX often outputs ungrammatical recognition results because the baseline SPHINX system uses very simple word-pair or bigram language models. For applications of spoken-language processing, sentence accuracy is also important. We have incorporated LR parsing into SPHINX to provide more global grammatical constraints. However, when using a tight grammar in a left-to-right continuous speech recognizer, once things begin to go wrong, they tend to go completely wrong. We also propose a solution to this problem by combining a context-free grammar with a bigram grammar.