Position Information for Language Modeling in Speech Recognition

This paper considers word position information for language modeling. For organized documents, such as technical papers or news reports, the composition and the word usage of articles of the same style are usually similar. Therefore, the documents can be separated into partitions consisting of identical rhetoric or topic styles by the literary structures, e.g., introductory remarks, related studies or events, elucidations of methodology or affairs, conclusions of the articles, and references, or footnotes of reporters. In this paper, we explore word position information and then propose two position- dependent language models for speech recognition. The structures and characteristics of these position-dependent language models were extensively investigated, while its performance was analyzed and verified by comparing it with the existing n-gram, mixture- and topic-based language models. The large vocabulary continuous speech recognition (LVCSR) experiments were conducted on the broadcast news transcription task. The preliminary results seem to indicate that the proposed position-dependent models are comparable to the mixture- and topic-based models.