The concept of an embodied interaction robots system based on speech is proposed for assisting human interaction by generating the motions of robots coherentlyrelated to speech. For an essential human interaction model which estimates a listener’s nodding on the basis of a speaker’s burst-pause of speech, the validity of a moving-average (MA) model is demonstrated in the robustness of prediction. Then, the prototype of the proposed system which consists of 2 interaction robots with both functions of speaker and listener is developed in which two models are incorporated: one is a listener’s reaction model in which nodding, blinking and the motions of head, arms and body are estimated by the MA model of the burstpause of speech to the nodding: the other is a speaker’s bodily reaction model in which the motions of head, arms and body are estimated by its own MA model of the burst-pause of speech to the head motion. By the sensory evaluation and behavioral analysis in humanrobot interaction, the effectiveness of the robot’s bodily interaction is demonstrated. The effectiveness of the embodied interaction robots system is also confirmed in remote communication.