Audio source separation by source localization and spatial beamforming with binaural model

Summary form only given. This paper presents a method of localization based audio source separation from binaural mixtures with adaptive beamforming. A new source localization technique is proposed to spatially localize the sources with two degrees of freedom (azimuth and elevation). A head related transfer function is used to synthesize the binaural mixture that introduces the interaural differences to the source signals. The interaural cues are technically employed to estimate the directions of arrival of the sound sources. A data independent minimum variance beamfilter is used to segregate the sources sequentially to overcome the underdetermined condition (more sources than sensors). The experimental results using the proposed algorithm show a noticeable efficiency in audio source separation. We have also presented the separation performance by a human hearing experiment, comparing the original and separated signals.