Using rules to improve letter to sound conversion of names

This paper presents an investigation of the use of context sensitive rewrite rules for improving the performance of data driven letter to sound conversion, concentrating on the specific case of British names. Taking a practical point of view, emphasis is put on reduction of the worst phonetization errors, and on improving the maintainability of the system -helping in database cultivation, and allowing an easier way of correcting errors. The rules are shown to be useful in cleaning the training database, improving the letter to phoneme alignment, and helping control the structure of the decision trees. Evaluation has shown that the suggested method outperforms the pure data driven method both in objective and subjective tests, on British and American name tasks.