Abstract
With Taiwanese text written in LangGeh orthography, we study the automatic phonetic annotations of the text. Compared to the baseline case that uses Daiim phonetic dictionary (Chiang 2002) and forward maximal matching, we study some possible improvements using various information extracted from corpus: using high frequency phones of single characters as well as multi-syllabic words, positional information of a character in a LangGeh phrase, and additional phonetic dictionary extracted from parallel corpus. Due to limitation of corpus size, only high frequency phones of characters exhibits significant improvement in our experiments.