DAFX: Digital Audio Effects, Second Edition

14.2 Binaural Source Separation

In binaural separation, the various sources are segregated according to spatial cues extracted from the signals available at both ears. This is the case for signals transduced by the microphones of a two-ear hearing-aid system or of the signals available at the microphones in the artificial ears of a dummy head. The type of apparatus is irrelevant, as long as human-like head shadowing is present in-between the microphones. A common strategy is to first detect and discern sources based on the DOA of waves, which is the object of Section 14.2.1, and then to build suitable time-frequency masks in a two-channel STFT representation of the binaural signal in order to demix them. Each mask, which can be binary or continuous valued with arbitrary values in [0, 1], coarsely represents the time-frequency profile of a given source. Due to energy leakage in finite sliding window time-frequency analysis, the masks are bound to cover broader bands than the ideal time-frequency tracks. The estimation of proper masks is the subject of Section 14.2.2. The masks are multiplicatively applied to both STFT channels and the transforms inverted. The ideal outcome is a set of binaural signals, each representing the sound of a single source spatially placed at the corresponding DOA.

14.2.1 Binaural Localization

An important aspect of binaural localization is the estimation of the DOA in terms of azimuth and elevation. Together with the range information (distance from the ...