Sound Source Distance Learning Based on Binaural Signals

A learning approach for estimating sound source distance based on binaural signals is presented. The frequency-dependent coherence between the left and right ear signals is used as the distance cue. The distance estimation is based on pre-calculated coherence profiles and an energy-weighted likelihood function. The system is evaluated with different speech samples. The accuracy is best in the frontal direction and poorer in the sides, due to shadowing of the head. However, the system shows promise for scenarios where the sound source location is restricted and could be integrated with a two-channel azimuth localization system.