Reversible letter-to-sound/sound-to-letter generation based on parsing word morpology

Abstract

This paper describes a bi-directional letter/sound generation system based on a strategy combining data-driven techniques with a rule-based formalism. Our approach provides a hierarchical analysis of a word, including stress pattern, morphology and syllabification. Generation is achieved by a probabilistic parsing technique, where probabilities are trained from a parsed lexicon. Our training and testing corpora consisted of spellings and pronunciations for the high frequency portion of the Brown Corpus (10,000 words). The phonetic labels are augmented with markers indicating morphology and stress. We will report on two distinct grammars representing a historical perspective. Our early work with the first grammar inspired us to modify the grammar formalism, leading to greater constraint with fewer rules. We evaluated our performance on letter-to-sound generation in terms of whole word accuracy as well as phoneme accuracy. For the unseen test set, we achieved a word accuracy of 69.3% and a phoneme accuracy of 91.7% using a set of 52 distinct phonemes. While this paper focuses on letter-to-sound generation, our system is also capable of generation in the reverse direction, as reported elsewhere (Meng et al., 1994a). We believe that our formalism will be especially applicable for entering unknown words orally into a recognition system.