Binaural Sound Source Distance Learning in Rooms

Abstract

A method for learning the distance of a sound source in a room is presented. The proposed method is based on short-time magnitude-squared coherence between the two channels of a binaural signal. Based on white noise as the training data, a coherence profile is obtained at each desired position in the room. These profiles can then be used to identify the most likely distance of a speech signal in the same room. The proposed approach is compared to a previous method for learning the position of a sound source. The results indicate that the both methods are able to identify the distance of a speech sound source correctly in a grid with 0.5-m spacing in most cases, when the orientation of the listener is 0deg , 30deg , 60deg , 90deg , or 180deg on the horizontal plane.