Kernel method for speech source activity detection in multi-modal signals
David Dov, Ronen Talmon, Israel Cohen · 2016
We consider a problem setup, in which a desired speech source is measured by a microphone and by a video camera in an interfering environment. We assume that the interfering sources in the audio signal are independent of the interfering sources in the video signal (e.g., the video signal does not capture the interfering speakers). Our objective in this paper is to detect the activity of the desired source. To address this problem, we take a kernel based geometric approach for obtaining a representation of the measured signal, in which the effect of the interfering sources is reduced. Based on this representation, we devise a measure for the activity of the desired source; experimental results demonstrate its superiority compared to competing methods in the detection of speech signals in the presence of different challenging types of interferences, including interfering speakers in the audio signal.