IMPROVED ESTIMATION OF PROBABILITIES IN PRONUNCIATION BY ANALOGY

Pronunciation by Analogy is a method for generating phonetic transcriptions for previously unseen written words based on matching substrings of known words and their pronunciations. The method inherently generates several candidate pronunciations and a multitude of heuristics have been proposed for choosing the best one. In [1], a theoretically justified probabilistic approach for scoring the pronunciations was proposed, with performance on par with the best heuristic methods. However, a certain ad hoc modification-a fractional power applied to the estimated probabilities of the substring pronunciations-was also found to improve performance. In this article, we give an explanation for this unexpected improvement. We show that the fractional power in fact improves the estimates of the candidate pronunciation probabilities. This also gives an indirect explanation of the good performance of the current best heuristic proposed in [2].