Statistically trained orthographic to sound

Many languages have a non-obvious, but not unrelated, relationship between orthography and pronunciation. Traditional methods for automatic conversion from letters to phones involve hand-crafted letter-to-sound rules, but these require care and expertise to develop. This paper presents a letter-to-sound rule system for Thai, that is trained automatically from lexicons. A statistical model, decision trees, is used to predict phones from letters. Letters mappping to multi-phones are used to solve the problem of implicit vowels and final consonants propagation and preand post-processing techniques are used to handle the inversion of initial consonants and vowels. For tone prediction, hand-crafted rules are used instead since there is no ambiguation if the phonological composition is known. Combining the n-gram of phone model with the decision trees, we can achieve 68.76% word accuracy which is better than 65.15% word accuracy in the rule-based approach.