A High Accuracy Approach for Word-Phoneme Translation Using Neural Networks

This paper presents a high accuracy approach for word-phonetics translation from a given lexicon for use in pronouncing out-of-vocabulary words and as a method for lexicon compression. We introduce grapheme, not letter to produce reasonable alignment of graphemes to phonemes. Neural networks model are trained on the aligned entries and used to predict the pronunciation of new words. For the CMU lexicon we have tested, our models have a word accuracy of 78.33%, higher than the published approaches