DOA estimation of speech signals using semi-blind source separation techniques

In this paper we investigate the application of complex independent component analysis (ICA) to the direction of arrival (DOA) estimation problem of wideband signals. The ICA based technique is semi-blind in the sense that the structure of the array is known to be uniform and linear (ULA). We show that when the array is ULA the mixing matrix is forced to have the structure imposed by the directivity vectors of the microphone array. ICA is applied to the spectral bins having the higher SNRs. The DOAs are derived from the histogram and clustering of the angle of arrivals of all high SNR spectral bins. The effectiveness of the approach is evaluated on speech signals and is compared against a variety of wideband DOA estimation techniques based on second order statistics. Finally, a toolbox for DOA estimation of wideband sources is made available at: http://slt.wcl.ee.upatras.gr/potamitis/DOA_ICA.zip