A robot quizmaster that can localize, separate, and recognize simultaneous utterances for a fastest-voice-first quiz game

This paper presents an interactive humanoid robot that can moderate a multi-player fastest-voice-first-type quiz game by leveraging state-of-the-art robot audition techniques such as sound source localization and separation and speech recognition. In this game, a player who says 'Yes' first gets a right to answer a question, and players are allowed to barge in a questionary utterance of the quizmaster. The robot needs to identify which player says 'Yes' first, even if multiple players respond at almost exactly the same time, and must judge the correctness of the answer given by the player. To enable natural human-robot interaction, we believe that the robot should use its own microphones (i.e., ears) embedded in the head, rather than having pin microphones attached to individual players. In this paper we use a robot audition system called HARK for separating the mixture of audio signals recorded by the ears into multiple source signals (i.e., almost the simultaneous utterances of 'Yes' and the questionary utterance) and estimating the direction of each source. To judge the correctness of an answer, we use a speech recognizer called Julius. Experimental results showed that our robot can correctly identify which player spoke first when the players' utterances differed by 60 msec.