Grapheme-to-phoneme conversion based on TBL algorithm in Mandarin TTS system

Grapheme-to-phoneme (G2P) conversion is an important component in a Text-to-Speech (TTS) system. The difficulty in Chinese G2P conversion is to pick out one correct pronunciation from several candidates according to the context information. By evaluating the distribution of polyphones in a corpus with manually corrected pinyin transcriptions, this paper pointed out that the overall error rate of G2P conversion was greatly decreased after processing 78 key polyphones. This paper proposed a transformation-based error-driven learning (TBL) algorithm to solve G2P conversion for polyphones. The correct rates of G2P for polyphones, which originally have high accuracy or low accuracy, are both improved. Besides, two additional experiments show that the capacity of the TBL algorithm has great relationships with initial status and TBL algorithm is more suitable than decision tree to solve polyphones’ G2P problem. Keyword: Grapheme-to-Phoneme (G2P); TransformationBased Learning (TBL); Decision Tree (DT); polyphone