SPHERICAL CONVOLUTIONAL RECURRENT NEURAL NETWORK FOR REAL-TIME SOUND SOURCE TRACKING

Neural networks have been widely applied in Direction-of-Arrival (DOA) estimation and source tracking system. In this paper, we introduce the Spherical Convolutional Recurrent Neural Network which utilizes laplacian graph-based spherical convolution using (SRP-PHAT) power maps for real-time robust sound source DOA estimation and tracking application. This proposed method achieves a similar performance with the state-of-the-art 3D convolutional neural networks method and declines the processing time by 88.6%, the parameter number by 85.5%, training memory usage by 54.0% respectively by a fairly shallow structure which demonstrates the effectiveness and efficiency of the proposed method.