Speaker detection using the timing structure of lip motion and sound

Yu Horii, Hiroaki Kawashima, Takashi Matsuyama

Abstract: In this paper, we propose a novel approach to speaker detection by an integration of audio-visual information using the cue of timing structure. We first extract feature sequences of lip motion and sound, and segment each of them into temporal intervals. Then, we construct a cross-media timing-structure model of human speech by learning the temporal relations of overlapping intervals. Based on the learned model, we realize speaker detection by evaluating the timing structure of the observed video and audio. Our experimental result shows the effectiveness of using temporal relations of intervals for speaker detection.