Sound classification and localization in service robots with attention mechanisms

Human-machine interaction is calling for a sophisticated understanding of subjects’ behavior performed by smartphones, home automation and entertainment devices, and many service robots. During an interaction with human beings in their environment, a service robot has to be capable to perceive and process visual and sound information of the scene that he observes. To capture salient elements in such different signals many semi-supervised deep learning methods have been proposed. In this article, it is proposed a new convolutional neural network, endowed with a mechanism of attention in order not only to classify, but also to localize temporally a sound event, and in a semi-supervised way.