Time frequency source separation and direction of arrival estimation in a 3D soundscape environment

The DUET (Degenerative Unmixing and Estimation Technique) algorithm has been shown to be an effective method for the separation of multiple sources from a stereo mixture. We present a time–frequency masking technique based on DUET for the blind separation of N sources from a four channel B format audio mixture. This method is applicable where sources that are both radially sparse in three dimensions, and exhibit approximate W-disjoint orthogonality, i.e. where the signals are approximately sparse in the time–frequency domain. Using the B-format mixture, we generate a three dimensional power-weighted geodesic histogram and show the peak locations correspond to the direction of arrival of the sources. These peaks are then used to generate a time–frequency mask to separate the sources from the w-channel of the B-format mixture. Experimental separation results using speech signals are presented.