Pronunciation Modelling and Lexical Adaptation in Mid-size Vocabulary Asr

A computational-phonological method is presented to automatically adapt the phone transcriptions in a lexicon to improve ASR performance in a number of mid-size recognition tasks. The lexical adaptation approach is based on supervised phoneme loops using cd-HMM segments to find alternatives for the transcriptions, and can be considered as a counterpart of the K-means algorithm but on symbolic level. The word error rate in a limited task (digit string recognition) with dialect speakers is shown to drop by 20-25 percent relative, starting from non-dialect digit transcriptions. Since the method is computationally involving, it is only feasible for relatively small tasks.