Robust Indoor Robotic Auditory Tracking with Icosahedral Residual CNN

Robust indoor robotic auditory tracking remains challenging due to the interaction between reverberation and diffuse noise, etc. In this paper, we present an icosahedral residual network to localize a single sound source, which is captured with a robot head microphone array. The Icosahedral Convolutional Neural Network has the advantages of low computational cost and reasonable approximation to the sphere. In addition, to improve the system's robustness, icosahedral SRP-SVD power maps are used as input features. The experimental result of the actual recordings in the LOCATA dataset performs accurately on tracking of a single sound source. Our system achieves significantly improved localization performance even in highly reverberant indoor environments.