Robot audition for dynamic environments

摘要

This paper addresses robot audition for dynamic environments, where speakers and/or a robot is moving within a dynamically-changing acoustic environment. Robot Audition studied so far assumed only stationary human-robot interaction scenes, and thus they have difficulties in coping with such dynamic environments. We recently developed new techniques for a robot to listen to several things simultaneously using its own ears even in dynamic environments; MUltiple SIgnal Classification based on Generalized Eigen-Value Decomposition (GEVD-MUSIC), Geometrically constrained High-order Decorrelation based Source Separation with Adaptive Step-size control (GHDSS-AS), Histogram-based Recursive Level Estimation (HRLE), and Template-based Ego Noise Suppression (TENS). GEVD-MUSIC provides noise-robust sound source localization. GHDSS-AS is a new sound source separation method which quickly adapts its sound source separation parameters to dynamic changes. HRLE is a practical post-filtering method with a small number of parameters. ENS estimates the motor noise of the robot by using templates recorded in advance and eliminates it. These methods are implemented as modules for our open-source robot audition software HARK to be easily integrated. We show that each of these methods and their combinations are effective to cope with dynamic environments through off-line experiments and on-line real-time demonstrations.