Independent Vector Analysis Assisted Adaptive Beamfomring for Speech Source Separation with an Acoustic Vector Sensor

Acoustic vector sensor (AVS), as a compact sensor with the capability of forming a frequency-invariant spatial beampattern over the 3D space, has potential in source separation. A straightforward way to achieve source separation with AVS is through adaptive beamforming. Such a method requires the direction-of-arrival (DOA) information, which is challenging to estimate accurately in reverberant environments. To circumvent this issue, we present a framework jointly implementing adaptive beamforming and independent vector analysis (IVA). Different from the conventional beamforming, the presented method only require rough DOA estimation for initialization. It iteratively refines the estimates of source DOA and signal statistics. The proposed method has great advantages of improving source separation performance and enhancing DOA estimation accuracy. Simulations demonstrate the properties of the developed method.