Recent work at the Centre for Speech Technology Research (CSTR, The University of Edinburgh) has developed a dialect/accent-independent lexicon for speech synthesis. The main purpose of this lexicon is to by-pass the problems and cost of writing a new lexicon for every new dialect/accent needed for synthesis.
The problem of how to deal with pronunciation variation for automatic speech recognition is currently a hot topic. It is clear that there is variation in pronunciation (at least, across speakers) which cannot be handled solely by the acoustic models. Therefore, mul- tiple pronunciations for some words must be included in the lexicon. This can improve recognition accuracy, but if too many variants are included, accuracy actually decreases. Since variation within a single speaker is relatively small we propose a novel method using speaker-specific lexica for ASR, which include only those pronunciation variants appropriate for a single speaker's accent. The main aim of this project is therefore to use the Keyword Lexicon to generate such speaker-specific lexica, and use these lexica in a standard HMM-based recognition system. We will use a standard speech recognition benchmark task (WSJCAM0 - a British English version of the Wall Street Journal corpus) to evaluate our new method.