Abstract
We have developed a pen-based Japanese character input system for use by blind people (particularly acquired blind person). Blind persons that are novices to system use can input Japanese characters without prior training, and information displayed on the screen is given to the user via the system's built-in voice synthesizer. The system can recognize 3, 216 characters including JIS Level 1 characters. However, when the system was applied to an E-mail system, many characters having a small number of strokes (such as Hiragana and Katakana) were contained in the mail text, causing a decrease in the recognition rate as the features used for recognizing a character were fewer. This paper describes an improvement made in recognition accuracy by combining the voting method with our previous method and two other conventional recognition methods. Character recognition of a phrase unit was also extended using a tri-gram probable language model. By applying these additional improvement methods, the recognition accuracy has been improved to 97.7% from 66.2%, and we confirmed the validity of the proposed method.
