Sound source detection threshold estimation using negative coherent power

This paper introduces a method for automatic detection and location of sound sources using distributed microphones in conjunction with a steered response power (SRP) algorithm. In many practical sound source location applications, sources must be automatically detected for computer driven analyses of the auditory scene. The method introduced in this paper takes advantage of the near-symmetric noise distribution of the coherent power resulting from the cross-correlations between microphone pair signals; autocorrelation components, which are included in the conventional SRP image, are removed resulting in positive and negative image values (referred to as coherent power). Thresholds applied to the positive coherent power values are estimated adaptively using local neighborhoods of negative coherent power values to achieve a constant false alarm rate (CFAR). In order to ensure sufficient symmetry, this paper derives a statistic based on the microphone geometry and the field of view (FOV) that is useful for determining source frequency ranges that result in effective zero-mean symmetric distributions necessary for an accurate CFAR threshold estimation. Experimental results for several microphone geometries are presented to assess the CFAR threshold accuracy. False-alarm probabilities are tested over a range from 10² and 10⁶ and error computed in terms of the experimentally assessed FA probability to the specified FA probability. Results show good agreement between the microphone geometry statistic and the relative performance of the CFAR thresholds. The best performing geometry achieved a false-alarm rate error of 16% (at 10⁶) when a Weibull distribution model was used for CFAR threshold estimation.