A weighted MVDR beamformer based on SVM learning for sound source localization

•A WMVDR beamformer for sound source localization in a reverberant room is proposed.•The weighted coefficients are modeled by a SVM classifier.•The skewness measure of marginal distributions is proposed as input feature.•Classify the narrowband power maps into constructively and disruptively contributing. A weighted minimum variance distortionless response (WMVDR) algorithm for near-field sound localization in a reverberant environment is presented. The steered response power computation of the WMVDR is based on a machine learning component which improves the incoherent frequency fusion of the narrowband power maps. A support vector machine (SVM) classifier is adopted to select the components of the fusion. The skewness measure of the narrowband power map marginal distribution is showed to be an effective feature for the supervised learning of the power map selection. Experiments with both simulated and real data demonstrate the improvement of the WMVDR beamformer localization accuracy with respect to other state-of-the-art techniques.