Pitch-Cluster-Map Based Daily Sound Recognition for Mobile Robot Audition

Abstract

This paper presents a sound identification method for a mobile robot in home and office environments. We propose a short-term sound recognition method using Pitch-Cluster-Maps (PCMs) sound database (DB) based on a Vector Quantization approach. A binarized frequency spectrum is used to generate PCMs codebook, which describes a variety of sound sources, not only voice, from short-term sound input. PCMs sound identification requires several tens of milliseconds of sound input, and is suitable for mobile robot applications in which conditions are continuously and dynamically changing. We implemented this in mobile robot audition system using a 32-channel microphone array. Robot noise reduction and sound source tracking using our proposal are applied to robot audition system, and we evaluate daily sound recognition performance for separated sound sources from a moving robot.