Teaching a robot to hear: a real-time on-board sound classication system for a humanoid robot

Abstract

We present an approach for detecting, classifying and recognising novel non-verbal sounds on an Aldebaran Nao humanoid robot. Our method allows the robot to detect novel sounds, classify these sounds, and then recognise future instances. To learn the names of sounds, and whether each sound is relevant to the robot, a natural speech-based interaction occurs between the robot and a human partner in which the robot seeks advice when a novel sound is heard. We test and demonstrate our system via an interactive human-robot game in which a person interacting with the robot can teach the robot via speech the names of novel sounds, and then test the robot’s auditory classication and recognition capabilities by providing further examples of both novel sounds and sounds heard previously by the robot. The implementation details of our acoustic sound recognition system are presented, together with empirical results describing the system’s level of performance.