Improving automatic phonetic transcription of spontaneous speech through variant-based pronunciation variation modelling

In this paper we present an experiment aimed at improving automatic phonetic transcription of Dutch spontaneous speech through a variant-based method of pronunciation variation modelling. For spontaneous speech, the literature does not always provide enough rules to describe its characteristic phonological processes. Therefore, other methods should be applied to model pronunciation variation for automatic phonetic transcription. We show that a large amount of manually transcribed phonetic data is an extremely useful source for collecting pronunciation variants and their prior probabilities. From the results we can conclude that the adopted method is indeed suitable for improving automatic transcription of spontaneous speech, and that further improvements can be obtained by combining this method with rule-based methods of pronunciation variation modelling