Spherical Convolutional Recurrent Neural Network for Real-Time Sound Source Tracking
Tianle Zhong, Israel Mendoza Velázquez, Yi Ren, Héctor Manuel Pérez Meana, Yoichi Haneda
Abstract
Neural networks have been widely applied in direction-of-arrival (DOA) estimation and source tracking systems. In this paper, we introduce a spherical convolutional recurrent neural network that utilizes Deepsphere, a graph-based spherical convolutional neural network, employing the steered response power with phase transform (SRP-PHAT) power maps as input features for real-time robust sound source DOA estimation and tracking applications. The proposed method achieves a performance similar to that of state-of-the-art 3D convolutional neural networks (3D-CNNs) method and reduces the processing time by 88.6%, the parameter count by 85.5%, and the training memory usage by 54.0% respectively. The shallow structure of proposed network demonstrates effectiveness and efficiency.
BibTeX
@inproceedings{icassp2022_sphericalconvolu,
title = {Spherical Convolutional Recurrent Neural Network for Real-Time Sound Source Tracking},
author = {Tianle Zhong and Israel Mendoza Velázquez and Yi Ren and Héctor Manuel Pérez Meana and Yoichi Haneda},
booktitle = {ICASSP 2022},
year = {2022}
}