Deep Neural Networks for Multiple Speaker Detection and Localization
share
We propose to use neural networks (NNs) for simultaneous detection and localization of multiple sound sources in Human-Robot Interaction (HRI). Unlike conventional signal processing techniques, NN-based Sound Source Localization (SSL) methods are relatively straightforward and require no or fewer assumptions that hardly hold in real HRI scenarios. Previously, NN-based methods have been successfully applied to single SSL problems, which do not extend to multiple sources in terms of detection and localization. In this paper, we thus propose a likelihood-based encoding of the network output, which naturally allows the detection of an arbitrary number of sources. In addition, we investigate the use of sub-band cross-correlation information as features for better localization in sound mixtures, as well as three different NN architectures based on different processing motivations. Experiments on real data recorded from the robot show that our NN-based methods significantly outperform the popular spatial spectrum-based approaches.
READ FULL TEXT