Statistical and Linguistic Knowledge Based Speech Recognition System: Language Acquisition Device for Machines

摘要

Today’s speech recognizers use very little knowledge of what language really is. They treat a sentence as if it would be generated by a random process and pay little or no attention to its linguistic structure. If recognizers knew about the rules of grammar, they would potentially make less recognition errors. Highly linguistically motivated grammars that are able to capture the deeper structure of language have evolved from the natural language processing community during the last few years. However, the speech recognition community mainly applies models which disregard that structure or applies very coarse probabilistic grammars. This paper aims at bridging the gap between statistical language models and elaborate linguistic grammars. Firstly an analysis of the need to integrate the conventional Statistical Language Models with the modern Linguistic Knowledge based language models is made, thereby justifying the Statistical and Linguistic Knowledge based Speech Recognition System which is asymptotically error free.