We present the weighted minimum variance distortionless response (WMVDR), which is a steered response power (SRP) algorithm, for near-field speaker localization in a reverberant environment. The proposed WMVDR is based on a machine learning approach for computing the incoherent frequency fusion of narrowband power maps. We adopt a radial basis function network (RBFN) classifier for the estimation of the weighting coefficients, and a marginal distribution of narrowband power map as feature for the supervised training operation. Simulations demonstrate the effectiveness of the proposed approach in different conditions.
Cite as: Salvati, D., Drioli, C., Foresti, G.L. (2015) Frequency map selection using a RBFN-based classifier in the MVDR beamformer for speaker localization in reverberant rooms. Proc. Interspeech 2015, 3298-3301, doi: 10.21437/Interspeech.2015-664
@inproceedings{salvati15_interspeech,
author={Daniele Salvati and Carlo Drioli and Gian Luca Foresti},
title={{Frequency map selection using a RBFN-based classifier in the MVDR beamformer for speaker localization in reverberant rooms}},
year=2015,
booktitle={Proc. Interspeech 2015},
pages={3298--3301},
doi={10.21437/Interspeech.2015-664}
}