Mean shift based direction finding and separation of multiple sources using microphone array in the presence of spatial aliasing

This paper proposes a mean shift based method of blind source separation as well as source direction estimation through time-frequency (T-F) decomposition. The basic cue of T-F cell adopted here is the wave propagation vector and the method is applicable to arbitrary array configuration in 3-D space. In addition, it also applicable to wide extent array, thus the method cope with spatial aliasing problem. In order to solve such aliasing problem causing ambiguity of phase, norm normalization property of propagation direction vector is employed. By applying mean shift for both mode search and clustering of probability density estimation, efficient separation procedure is realized as a binary masking in T-F domain. The proposed approach is implemented in four microphones with wide sensor spacing where spatial aliasing may occur. Efficient separation results even for multiple sources are demonstrated.