Testing a Method for Modelling Pronunciation Variation

least part of the pronunciation variation that is encountered in the In this paper we describe a method for improving the performthose used previously with success in [3, 4]. ance of a continuous speech recognizer by modelling pronunIn this approach phonological rules are used to generate ciation variation. Although the results obtained with this method pronunciation variants, i.e. to expand the lexicon. The expanded are in line with those reported by other authors, the magnitude of lexicon can then be used during training, recognition (test) or the improvements is very small. In looking for possible both. During test the old test lexicon is simply replaced by the explanations for these results, we computed various sorts of new one, in order to make it possible to recognize pronunciation statistics about the material. Since these data proved to be very variants. During training the pronunciation variants can be used useful in understanding the effects of our method, they are to obtain new acoustic models as follows. discussed in this paper. Moreover, on the basis of these statistics ’ Forced recognition is carried out to determine which variant we discuss how the system can be improved in the future. is realized in the corpus. This way a new transcription of the