Towards Speech Acquisition in Natural Interaction on ASIMO

The standard approach for teaching robots to communicate via speech is by providing the structure, statistics, and semantics of speech through a supervised, offline learning process. This process imposes constraints like a high degree of specialization to certain, predefined tasks. The resulting system is very rigid and lacks the ability to acquire new skills (e.g. words and their semantics). In contrast to this, children acquire language through observation of adults’ speech and, more importantly, in interaction with them. As a result their speech capabilities are very flexible and can adapt to new situations. Our research target is therefore to build a system that can learn to acquire speech in interaction with humans. The interaction aspect requires a hardware platform that can engage in a natural communication with humans in real-world environments. For this purpose we employ our humanoid robot ASIMO (see Fig. 1). To provide the robot with human-like speech communication abilities we are working on several aspects of sound processing, scene representation, and learning that will be outlined in more detail in the next sections.