Using morphology and phoneme history to improve grapheme-to-phoneme conversion

In this study four statistical grapheme-to-phoneme (G2P) conversion methods for canonical German are compared. The G2P models differ in terms of usage of morphologic information and of phoneme history (left context) information. In order to evaluate our models we introduce two measures, namely mean normalized Levenshtein distance for classification accuracy and conditional relative entropy for validation of phonotactic smoothness. The results show that morphologic information significantly improves G2P conversion and together with phoneme history leads to a better approximation of the original phonotactics. Furthermore with the benefit of morphology our models significantly outperform two well established G2P systems.