Independent automatic segmentation by self-learning categorial pronunciation rules

The goal of this paper is to present a new method to automatically generate pronunciation rules for automatic segmentation of speech the German MAUSER system. MAUSER is an algorithm which generates pronunciation rules independently of any domain dependent training data either by clustering and statistically weighting self-learned rules according to a small set of phonological rules clustered by categories or by re-weighting “seen” phonological rules. By this method we are able to automatically segment cost-effectively large corpora of mainly unprompted speech.