Thai Grapheme-Phoneme Alignment: Many-to-Many Alignment with Discontinuous Patterns

Abstract

Grapheme-phoneme aligned data is crucial to the grapheme-to-phoneme conversion system. Although manual alignment is possible, the task is tedious and time-consuming. Therefore, unsupervised alignment algorithms are proposed to reduce this alignment cost. Several efficient algorithms rely on the assumption that patterns are continuous, but the assumption is not true for Thai. When applying these algorithms to Thai grapheme-to-phoneme alignment, some pre-processing steps for discontinuous patterns are necessary. We propose an algorithm to align Thai graphemes and phonemes which directly incorporates the discontinuous patterns. The experiments show that the precision of the proposed alignment algorithm substantially increases from the conventional alignment with only continuous patterns while the recall decreases from the original method. As a result, the proposed algorithm achieves similar F1 to the conventional algorithm.