Digital Signal Processing and System Theory

Talk Martin Heckmann

Language Acquisition Embedded into Tutor-Robot Interaction

Date: 28.06.2010, 17:15 h - 18:15 h
Room: Aquarium

Dr.-Ing. Martin Heckmann
Honda Research Institute Europe GmbH
Offenbach, Germany


talk_2010_martin_heckmannChildren acquire language to a large extend in the interaction with their caregivers. Inspired by this observation we develop computational models and artifacts for the acquisition of language in an interactive scenario. The artifact bootstraps its representations with little a priori knowledge and can be taught by a human tutor. In this presentation I will highlight different aspects of our research in this domain.

First I want to talk about the learning of speech features as well as word and sub-word units. As speech features we explore a set of hierarchical spectro-temporal features which are learned in an unsupervised fashion based on the observed speech data. I will demonstrate that these spectro-temporal features carry information complementary to purely spectral features and that this information talk_martin_heckmann_2010_honda_robot is beneficial for speech recognition. Furthermore, I will show how incremental word learning can be bootstrapped via the use of phone-like speech units emerging from an unsupervised clustering process and efficient alignment of a few training samples. Next I will introduce an auditory attention system enabling a robot to filter relevant from irrelevant acoustic events.

Finally, I will briefly highlight a system which learns associations between acoustic labels and visual representations in interaction with its tutor. The auditory processing is based on the building blocks described above and the whole system is implemented on the Honda humanoid robot.