Evaluating Metrical Phonology - a Computational- Empirical Approach

This study aims at providing an empirical basis for the evaluation of predictions made by metrical phonology. The predictions are compared to perceptual syllable prominence annotated in a database of German read speech. An evaluation baseline was defined on the correlation between prominence ratings for different speakers. Then, different sets of rules for prominence prediction were tested and evaluated. The results indicated that a correct prediction of word accent is most crucial for good results. Another outcome was, that a prediction based on lexical class is more successful than a purely syntactic approach. Some issues concerning the status of euphony rules in German are discussed.