Reversible sound-to-letter/letter-to-sound modeling based on syllable structure

Abstract

This paper describes a new grapheme-tophoneme framework, based on a combination of formal linguistic and statistical methods. A context-free grammar is used to parse words into their underlying syllable structure, and a set of subword "spellneme"units encoding both phonemic and graphemic information can be automatically derived from the parsed words. A statistical n-gram model can then be trained on a large lexicon of words represented in terms of these linguistically motivated subword units. The framework has potential applications in modeling unknown words and in linking spoken spellings with spoken pronunciations for fully automatic new-word acquisition via dialogue interaction. Results are reported on sound-to-letter experiments for the nouns in the Phonebook corpus.