Learning-Based Three Dimensional Sound Localization Using a Compact Non-Coplanar Array of Microphones

One of the various human sensory capabilities is to identify the direction of perceived sounds. The goal of this work is to study sound source localization in three dimensions using some of the most important cues the human uses. Having robotics as a major application, the approach involves a compact sensor structure that can be placed on a mobile platform. The objective is to estimate the relative sound source position in three dimensional space without imposing excessive restrictions on its spatio-temporal characteristics and the environment structure. Two types of features are considered, interaural time and level differences. Their relative effectiveness for localization is studied, as well as a practical way of using these complementary parameters. A two-stage procedure was used. In the training stage, sound samples are produced from points with known coordinates and then are stored. In the recognition stage, unknown sounds are processed by the trained system to estimate the 3D location of the sound source. Results from the experiments showed under ±3° in average angular error and less than ±20% in average radial distance error.