Revealing Phonological Similarities be

In this paper, we present an approach to automatically revealing phonological classes within historically related languages. A newly created bilingual German-Dutch pronunciation dictionary is used for learning phonological similarities between the onsets, nuclei and codas of these two languages via EM-based clustering. Our evaluation is twofold: we apply the models to predict from a German word the phonemes of a Dutch cognate. The results show that it is harder to predict the pronunciation of the nucleus and the coda than the onset. We also evaluate our approach qualitatively, finding meaningful classes caused by historical sound changes.