Unsupervised graphoneme alignment evaluation for grapheme-to-phoneme conversion on complex asian-language orthographies

摘要

Grapheme-to-Phoneme conversion (G2P) is a key component in text-to-speech and speech-to-text applications. Developing G2P for a new language needs grapheme-phoneme aligned data that is a time-consumed task. Previous works introduced graphoneme, m-to-n grapheme-phoneme sequence pair, as an alignment unit that can be applied to unsupervised alignment approach. Due to complex grapheme-phoneme alignment in Asian languages; for instance, inter-syllabic and intra-syllabic grapheme-phoneme cross-alignment and implicit tone assignment from orthographies, conventional contiguous character-based graphonemes might not be applicable. Thus, this paper investigates G2P performances using unsupervised graphoneme-based alignment techniques with varied lengths of graphonemes across Asian languages. We vary the lengths of graphonemes ranging from character level to supra-syllable level. To evaluate the G2P performance over Asian languages, we apply the alignments on four Asian grapheme-phoneme dictionaries including Mongolian, Lao, Thai and Myanmar. The evaluations of the unsupervised graphoneme-based alignments across the Asian languages yield high word accuracies of 79.18%, 81.63%, 91.28% and 95.13% for G2P applications in Myanmar, Thai, Lao and Mongolian, respectively.