Temporal Adaptation of Language Models

摘要: In this paper, the implications of recognising speech from news broadcasts that change on a daily basis are investigated perspective audio indexing. First, vocabulary coverage is found to be great importance. newspapers and it observed half total number unique words occur over time used only one day never again. The extent which vocabularies can adapted examined difficult obtain substantially increased using external sources time-dependent data. second implication time-varying data ability effectively adapt language model in recogniser so as best match being recognised. Language adaptation different methods combining domain dependent both fixed vocabularies. It improvements obtained ten consecutive shows radio programme Marketplace do not generally justify effort involved adapting models especially if baseline well trained.