Character recognition techniques are expected to be used for document input. Their recognition accuracy is however not always high for handwritten characters. Since the confirmation-correction input mode, used currently, becomes inefficient when there are many recognition errors, input methods relying solely on this mode are not suitable for document input when recognition accuracy is low. This paper proposes a document input method that selects the optimum input mode based on character recognition accuracy. This method automatically selects the optimum input mode using the relation between recognition accuracy and input efficiency and can increase overall input speed (the number of inputted characters per unit time) for any accuracy. Experimental results using handwritten documents verify that with this method input speed increases as a whole in comparison with input methods using a single input mode.