Using Latent Semantic Analysis to Aid Speech Recognition and Understanding

Generally, speech recognition engines can employ two different grammar methods, rule and dictation, to recognize an utterance. The purpose of these grammars is to constrain the search space in a way that anticipates the speaker’s utterance. The research described in this paper attempts to maintain the accuracy of a rule grammar without limiting the speaker to rigorous phraseology. Latent Semantic Analysis (LSA) is used to connect specific grammar rules with the meanings underlying matching phrases resulting in utterances being matched to knowledge base elements even though the exact phrase did not match any grammar rule. A separate knowledge base is used to dynamically add or remove grammar rules in the speech recognition engine as the conversation context changes. Finally, a learning technique is used to create new regular expressions based on utterances that matched semantically through LSA.