Using Natural-Language Knowledge Sources in Speech Recognition

High accuracy speech recognition requires a language model, to specify what word sequences are possible or at least likely. Standard n-gram language models for speech recognition ignore linguistic structures, but more more linguistically sophisticated language models are possible. Unification grammars are widely used in natural-language processing, and these can be compiled into into non-left-recursive context-free grammars that can then be used in real-time speech recognizers by dynamically expanding them into state-transition networks. A hybrid language model incorporating both a unification grammar and n-gram statistics has been shown to increase speech recognition accuracy. Probabilistic constext-free grammars and probabilistic unification grammars are also possible.