Rule based learning of word pronunciations from training corpora

Abstract

This paper describes a text-to-pronunciation system using transformation-based error-driven learning for speech-recognition purposes. Efforts have been made to make the system language independent, automatic, robust and able to generate multiple pronunciations. The learner proposes initial pronunciations for the words and finds transformations that bring the pronunciations closer to the correct pronunciations. The pronunciation generator works by applying the transformations to a similar initial pronunciation. A dynamic aligner is used for the necessary alignment of phonemes and graphemes. The pronunciations are scored using a weighed string edit distance. Optimizations were made to make the learner and the rule applier fast. The system achieves 73.9% exact word accuracy with multiple pronunciations, 82.3% word accuracy with one correct pronunciation, and 95.3% phoneme accuracy for English words. For proper names, it achieves 50.5% exact word accuracy, 69.2% word accuracy, and 92.0% phoneme accuracy, which outperforms the compared neural network approach.