Visual Based Reference for Enhanced Audio-Video Source Extraction

This paper addresses the problem of source extraction in a complex scene where only moving audio sources are present. An algorithm using a unique yet simple method avoiding higher-order statistics has been developed. The principle idea of the algorithm is to use a video camera array for locating a moving source whose position is used to isolate a noise reference, and thus allowing noise subtraction from the mixture based on the widely-known Widrow adaptive filtering method, that only uses second-order statistics. This adaptive approach provides an alternative to traditional methods particularly when there is need for a real time implementation.