Connecting concepts from vision and speech processing

Abstract

This paper addresses the problem of how to establish referential links between interpretations of speech and visual data. In order to get rid of erroneous, vague, or incomplete conceptual descriptions, we propose a probabilistic interaction scheme. The modelling of dependencies and the calculation of inferences are realized by using Bayesian networks. This interaction scheme provides a basis for disambiguation and error recovery. We implemented an interaction component in an assembly task environment. A robot constructor can be instructed by speech and pointing gestures in order to connect primitive component parts of a wooden toy construction kit. The system is evaluated on a test data set which consists of 448 spoken utterances from 16 speakers who name objects on 10 images from different scenes. First results show the effectiveness and robustness of the probabilistic approach.