Pronounce : a program for pronunciation by analogy

Humans can pronounce novel orthographically regular text strings such as pseudowords, like “peemers”, or words they have never seen before. How do they do this? Two hypotheses have been proposed to account for this ability. According to one view, pronunciation-by-rule, pseudowords are pronounced by a rule-based phonological process in which the pronunciation for a pseudoword is generated from its spelling by the use of a complex set of spelling-to-sound rules. According to the alternative view, pronunciation-by-analogy, pseudowords are pronounced by analogy to known words which are similar in spelling. Although the pronunciation-by-analogy approach is psychologically plausible, it is not clear that it is computationally feasible. Pronunciation-by-analogy depends on the degree to which orthographic consistency in the spelling patterns of words is related to phonotactic consistency in pronouncing those words. To investigate this theoretical issue, we developed a computer program called PRONOUNCE that automatically generates a set of rank-ordered pronunciations, in the form of a sequence of phonetic segments, using pronunciation-by-analogy with a lexicon of approximately 20 000 words based on Webster's Pocket Dictionary. PRONOUNCE examines every word in the lexicon and builds a pronunciation lattice structure using the phonetic representations of the words that match the input string. In this pronunciation lattice, each node represents a possible phoneme to be used at a particular position in the pronunciation, and each path through the lattice represents a possible pronunciation. At this time, PRONOUNCE performs reasonably well, generally producing pronunciations that agree with those given by native speakers of English. PRONOUNCE was tested on a set of 70 short pseudowords and was found to disagree with human subjects on only 9% of the pseudowords. These results suggest that pronunciation-by-analogy is indeed computationally feasible. Furthermore, the limited success of PRONOUNCE suggests a new approach to spelling-to-sound conversion for text-to-speech conversion systems.