A Hybrid Approach to Grapheme-Phoneme Conversion

We present a simple and effective approach to the task of grapheme-tophoneme conversion based on a set of manually edited grapheme-phoneme mappings which drives not only the alignment of words and corresponding pronunciations, but also the segmentation of words during model training and application, respectively. The actual conversion is performed with the help of a conditional random field model, after which a language model selects the most likely string of grapheme-phoneme segment pairs from the set of hypotheses. We evaluate our approach by comparing it to a state-ofthe-art joint sequence model with respect to two different datasets of contemporary German and one of contemporary English.