Default-and-refinement approach to pronunciation prediction

Abstract

The authors define a novel g-to-p prediction algorithm that utilises the concept of a 'default phoneme': a grapheme which is realised as a specific phoneme significantly more often than as any other phoneme. They found that this approach results in an algorithm that performs well across a range from very small to large data sets. The authors evaluated the algorithm on two benchmarked databases (Fonilex and NETtalk) and found highly competitive performance in asymptotic accuracy, initial learning speed, and model compactness.