A geometric interpretation of non-target-normalized maximum cross-channel correlation for vocal activity detection in meetings

Abstract

Vocal activity detection is an important technology for both automatic speech recognition and automatic speech understanding. In meetings, standard vocal activity detection algorithms have been shown to be ineffective, because participants typically vocalize for only a fraction of the recorded time and because, while they are not vocalizing, their channels are frequently dominated by crosstalk from other participants. In the present work, we review a particular type of normalization of maximum cross-channel correlation, a feature recently introduced to address the crosstalk problem. We derive a plausible geometric interpretation and show how the frame size affects performance.