Recognition Algorithms of the Target of the Utterances

this paper, we discuss the target recognition/identification of utterances, that is the problem to recognize the person to whom the speaker directed the utterance or the problem to identify if the utterance is directed to the recognizer or not. In the situation that the speakers are allowed to utter whatever they think freely, they tend to include some talks to someone other than the recognizer while they talk to the recognizer. They might include some talks to themselves also. Such utterances not directed to the recognizer make it difficult for the system to recognize the utterances of users. So it is very important to distinguish the utterances directed to the system from those not directed to the system. This problem is related to the rejection problem of the voice command input system. It is desired for such systems to accept only commands to the system and to reject non-command utterances such as soliloquies, talks to nearby persons, speeches of nearby persons. Usually, the system adopts the likelihood-based mechanisms to realize it. That means the system check if the likelihood functions is high enough to accept the utterance as a command. However, the absolute values of the likelihood functions are not reliable enough to decide if they are command or not because they are calculated only for the relative comparison to decide which command is suitable. Here, we attempt to apply two methods for the problems. One is the prosody-based approach. We usually change our mode of utterance according to the listeners. The typical example can be seen in mother tongue, she speaks to her baby softly and slowly than to the others. People tend to speak more courteously to their superiors than to their friends. Similarly, the utterance to the system might have special characters. ..