Automatic methods for lexical stress assignment and syllabification

Improvements in automatic lexical stress assignment and syllabification can increase the quality of text-to-speech synthesis as well as decrease the memory requirements for dictionaries. Several methods were evaluated. Machine-learning based methods are preferred since they easily adapt to multiple languages. For stress prediction, encouraging results were obtain by combining a decision tree approach with an algorithm that uses global (word level) statistical data derived from the training dictionary. For syllable boundary prediction, algorithms that learn syllable level statistics from the training dictionary perform very well, and can be implemented as a post-process after prediction of phoneme transcription and stress.