Video-informed approach for enhancing audio source separation through noise source suppression

This paper describes a method where an interference noise source within an audio source separation scenario is suppressed from a mixture. The principal idea of the proposed method is to use a video camera array for locating a interference noise source whose 3D position will be used to estimate a matrix of frequency responses (FRs) by linearly combining a series of previously known FRs. A filter is calculated to remove the contribution of the noise source from a convolutive mixture at each microphone, through the estimated FRs. The proposed method is assumed to implemented in a `block-wise´ manner in time domain and has been tested on mixtures created by impulse responses generated by the image method for small room acoustics.