Three-Dimensional Gradient-Based Tracking of Multiple Sound Sources

Abstract

This paper presents a gradient-based algorithm for tracking multiple sound sources in three-dimensional (3-D) space under reverberant conditions. The proposed method, called 3-D Gradient-based Multiple-source Tracking (3D-GMT), operates directly in the continuous spatial domain and estimates source trajectories via a frequency-constrained loss function. Unlike conventional grid-based approaches that suffer from quantization errors and require post-processing filters, 3D-GMT jointly refines all source locations in parallel, implicitly smoothing trajectories and reducing cumulative error. Simulation experiments in reverberant environments demonstrate the robustness and accuracy of 3D-GMT under two representative motion scenarios: straightline paths and complex spiral trajectories. The results indicate potential for applying the method to sound source tracking in reverberant environments with smooth, moderate-speed motion.