A Probabilistic Approach to Pronunciation by Analogy

TL;DR

This paper introduces a probabilistic scoring method for pronunciation by analogy, significantly improving accuracy in generating pronunciations for English words compared to previous approaches.

Contribution

It proposes a novel probabilistic scoring rule for PbA, achieving higher accuracy and demonstrating the benefits of a principled, data-driven approach over heuristic methods.

Findings

Achieved 66.21% accuracy on NETtalk corpus

Probabilistic scoring outperforms previous methods

Further improvements possible with combined approaches

Abstract

The relationship between written and spoken words is convoluted in languages with a deep orthography such as English and therefore it is difficult to devise explicit rules for generating the pronunciations for unseen words. Pronunciation by analogy (PbA) is a data-driven method of constructing pronunciations for novel words from concatenated segments of known words and their pronunciations. PbA performs relatively well with English and outperforms several other proposed methods. However, the best published word accuracy of 65.5% (for the 20,000 word NETtalk corpus) suggests there is much room for improvement in it. Previous PbA algorithms have used several different scoring strategies such as the product of the frequencies of the component pronunciations of the segments, or the number of different segmentations that yield the same pronunciation, and different combinations of these methods, to evaluate the candidate pronunciations. In this article, we instead propose to use a probabilistically justified scoring rule. We show that this principled approach alone yields better accuracy (66.21% for the NETtalk corpus) than any previously published PbA algorithm. Furthermore, combined with certain ad hoc modifications motivated by earlier algorithms, the performance climbs up to 66.6%, and further improvements are possible by combining this method with other methods.