Joint Detection and Localization of Multiple Speakers using a Probabilistic Interpretation of the Steered Response Power
Youssef Oualil, Mathew Magimai.-Doss, Friedrich Faubel, Dietrich Klakow · Infoscience (Ecole Polytechnique Fédérale de Lausanne) · 2012
Detection and localization of multiple speakers in a noisy and reverberant environment is a fundamental and difficult task. In the literature, steered response power (SRP) based techniques are typically used to accomplish this task which can be com-putationally intensive. Nonetheless, the localization of multiple speakers remains a challenging in practice. In this paper, we present a novel approach based on a probabilistic interpretation of the SRP. The proposed method replaces the discrete search techniques by proposing an approximate analytical form of the SRP, which can adequately detect and localize multiple speak-ers. In addition to reliable detection and localization, the poten-tial advantage of this approach is that it provides a probability density function (pdf) of the individual speaker positions rather than point estimates. Experiments on the AV16.3 corpus show the efficacy of the proposed approach.