DOA estimation of speech source in noisy environments with weighted spatial bispectrum correlation matrix
Wei Xue, Shan Liang, Wenju Liu · 2014
Although the high order statistics (HOS) has promising property against the Gaussian noise, there still lack effective ways to apply the HOS to DOA estimation of the speech source. In this paper, we propose a novel HOS based DOA estimation method for speech source in strong noise conditions. A “weighted spatial bispectrum correlation matrix (WSBCM)” is formulated, which contains the spatial correlation information of bispectrum phase differences. We then propose a new DOA estimator based on the eigenvalue analysis of the WS-BCM. Besides the theoretical advantage of the bispectrum against Gaussian noises, the redundant information in the bispectrum domain is also exploited to make the WSBCM noise robust. The WSBCM enables bispectrum weighting to select the speech units in the bispectrum, which further helps to improve the performance. Experimental results demonstrate that the proposed method outperforms existing algorithms in different kinds of noisy environments.