A dictionary‐intensive letter‐to‐sound program

We treat the letter‐to‐sound problem as that of deriving all regularly inflected forms of English words from a dictionary of base forms. A computer program must perform all lexical operations a human might have to perform in using an abridged desk dictionary. Of course this includes addition of “s.” “es,” “d,” “ed,” “er,” and “ing,” with their accompanying sound changes, deletions of mate “e,” and changes from “y”to “i.” The program must also handle other forms. Stress‐neutral endings like “less” and “ness” are quite simple. Four other types of endings, exemplified by “ation,” “ity,” “ian,” and “ency,” are more complex—reordering stress within the base word, and causing vowel shifts and consonant sound change. The problems with these endings are much like those for straight letter‐to‐sound rules, but with pronunciations of base form and ending as available information. With prefixes and compounds, the relative stress of constituent parts will depend on the intended part of speech for the target word, and on the properties of the parts. Final vowels of Greek prefixes may or may not reduce to schwa, depending on properties of the base forms they modify. With a programmed dictionary driver, we are currently finding ∼99.9% of words in the Brown corpus (excluding proper names and misspellings)—∼75% from exact dictionary entries, and the rest by derivation. We default to Mark Liberman's Name‐Say program for words not found or derived.