Abstract: A system comprises error microphones disposed in a predetermined pattern to capture audio signals including speech and one or more reference sensors to capture a reference noise signal from a noise source. A portion of the reference noise signal is removed from the captured audio signals to generate partially processed audio signals, which are paired to form signal pairs. For each signal pair, a modified cross-correlation vector, with effects of low frequency contents removed, is generated, then converted to a rotated angular domain cross-correlation vector based in part on a physical angle associated with locations of an associated pair of error microphones. The rotated angular domain cross-correlation vectors are summed, and a weighting vector is applied to the sum to identify direction information of a desired audio signal associated with the speech. The directional information is utilized to beamform the partially processed audio signals and output the desired audio signal.