Speaker
Description
Dynamic sound sources, such as footsteps or other moving sources, are common in real life but difficult to model because their characteristics and trajectory are often unknown. This paper presents a machine learning algorithm for sound field construction based on a temporal-frequency residual neural network. The algorithm is designed to capture how sound evolves across both time and frequency, effectively reconstructing the field even though the source trajectory and characteristics are unavailable. A key advantage of this algorithm is its ability to incorporate physical priors during training. By embedding acoustic propagation constraints as a regularization term, the neural network can better guide the reconstruction process toward physically consistent results. We tested the algorithm in simulated 2D and 3D scenarios; the results show that the proposed algorithm consistently outperforms the standard baseline in terms of reconstruction accuracy, while the physical prior significantly enhances stability under complex motion patterns.