Multiple active speaker localization based on audio-visual fusion in two stages

Abstract:

Localization of multiple active speakers in natural environments with only two microphones is a challenging problem. Reverberation degrades performance of speaker localization based exclusively on directional cues. The audio modality alone has problems with localization accuracy while the video modality alone has problems with false speaker activity detections. This paper presents an approach based on audiovisual fusion in two stages. In the first stage, speaker activity is detected based on the audio-visual fusion which can handle false lip movements. In the second stage, a Gaussian fusion method is proposed to integrate the estimates of both modalities. As a consequence, the localization accuracy and robustness compared to the audio/video modality alone is significantly increased. Experimental results in various scenarios confirmed the improved performance of the proposed system.