The present invention can be used to improve speech recognition accuracy, especially with regard to characters, words and the like which can correspond to a plurality of readings. The same person can be apt to maintain the same reading in the same conversation. For example, a person who pronounced "7" "shichi" is apt to pronounce "shichi" consistently in the conversation. By utilizing this tendency, recognition from the second time is executed after reducing a recognition probability corresponding to the reading, which is not used by the person in the first response of the conversation. In the case where a system repeats a recognition result by speech synthesis, the system repeats the recognition result corresponding to the reading of a speaker that is already recognized. For example, when the speaker pronounced "7" "shichi", the system pronounces "shichi" at the time of repetition.