In recent years, text-to-speech systems have been advanced, but they sometimes make mistakes when reading Japanese sentences including the kanji characters. Therefore, we developed an estimating system using a misreading ratio.

We then measured the misreading ratio of a commercial text-to-speech system, and analyzed the tendency of the misreadings. As a result, we found that the main reason for the misreadings of determing in the reading of homographs. Japanese homographs sometimes stand for words that have several meanings when writing, for example, '����' has two readings, 'kyoo' and 'kon-nichi'. Decision lists are used as a method to distinguish the homographs. We improved this method and achieved a high correct ratio.