Noise Robust Direction of Arrival Estimation for Speech Source With Weighted Bispectrum Spatial Correlation Matrix

One big challenge to the robust direction of arrival (DOA) estimation for the speech source is the environmental noise. In practical conditions, the noise can be undirected or emitted from a pointed source. In order to improve the reliability of DOA estimation in various adverse noisy conditions, we propose a novel DOA estimation method in this paper, and what lies in the core in the method is the “Weighted Bispectrum Spatial Correlation Matrix (WBSCM).” The bispectrum is a kind of higher order statistics (HOS) of a signal, and the WBSCM reflects the spatial correlation of the bispectrum phase differences (BPD) between different microphones. As the HOS of the Gaussian signal is theoretically zero, by formulating in the bispectrum domain, the proposed method has an inherent advantage against the Gaussian noise. Moreover, the BPD, which is embedded in the WBSCM, contains the redundant information related to the DOA of the speech source. This redundancy helps to improve the robustness in non-Gaussian noise conditions, especially for the directional interference scenarios. In addition, the WBSCM enables bispectrum weighting to select the speech units in the bispectrum, in order to highlight the effect of these units in the DOA estimation. Similar to the signal-to-noise estimation, a decision-directed method is proposed to compute the bispectrum weights. Finally, a new DOA estimator is proposed, which is based on the eigenvalue analysis of the WBSCM. We conduct experiments under various kinds of noisy environments, and the experimental results demonstrate the effectiveness of proposed method.