Phonological Parsing for Bi-directional Letter-to-Sound/Sound-to-Letter Generation

In this paper, we describe a reversible letter-to-sound/soundto-letter generation system based on an approach which combines a rule-based formalism with data-driven techniques. We adopt a probabilistic parsing strategy to provide a hierarchical lexical analysis of a word, including information such as morphology, stress, syllabification, phonemics and graphemics. Long-distance constraints are propagated by enforcing local constraints throughout the hierarchy. Our training and testing corpora are derived from the high-frequency portion of the Brown Corpus (10,000 words), augmented with markers indicating stress and word morphology. We evaluated our performance based on an unseen test set. The percentage of nonparsable words for letter-to-sound and sound-to-letter generation were 6% and 5% respectively. Of the remaining words our system achieved a word accuracy of 71.8~0 and a phoneme accuracy of 92.5% for letter-to-sound generation, and a word accuracy of 55.8% and letter accuracy of 89.4% for sound-to-letter generation. We also compared our hierarchical approach with an alternative, single-layer approach to demonstrate how the hierarchy provides a parsimonious description for English orthographic-phonological regularities, while simultaneously attaining competitive generation accuracy. I N T R O D U C T I O N This paper describes a trainable probabilistic system for reversible let ter-to-sound/sound-to-let ter generation. Sound-to-letter generation is a crucial aspect in the problem of automatic detection/incorporation of new words, which is in turn critical for the development of large vocabulary speech understanding systems. Moreover, letterto-sound generation will continue to be important for speech output , especially in applications such as reading machines. To successfully achieve our goal, several important issues must be addressed. First , what should be the inventory of linguistic or lexical units for describing English orthographic-phonological regularities? Second, how should these units be incorporated into the representation of English orthography and phonology? Third, what algorithms can be used to synthesize and analyze the spelling and pronunciation of an English word 1This research was supported by ARPA under Contract N0001489-J-1332, monitored through the Office of Naval Research, and a grant from Apple Computer Inc. in terms of these lexical units? These three issues will be addressed in detail in the following when we describe our approach and report on our system's performance for both letter-to-sound [1] and sound-to-letter generation [2]. The novel features of our approach include the reversibility of the combined parsing and generative processes, the abili ty to provide multiple output hypotheses, the capability of handling uncertainty in the input, as well as our t reatment of non-parsab!e words.