A method of dominant speech extraction is provided that includes acquiring a primary audio signal from a microphone and at least one additional audio signal from at least one additional microphone, wherein the acquired audio signals include speech and noise, decomposing each acquired audio signal into a low frequency sub-band signal and a high frequency sub-band signal, applying speech suppression beamforming to the low frequency sub-band signals to generate a reference channel having an estimate of noise in the low frequency sub-band signals, applying noise cancellation to the low frequency sub-band signal of the primary audio signal using the reference channel to generate a first signal having a low frequency estimate of the speech, applying noise suppression beamforming to the high frequency sub-band signals to generate a second signal having a high frequency estimate of the speech, and combining the first and second signals to generate a full-band audio signal.