Speaker
Description
Neural implicit fields are a powerful framework for sound field regression, but their application to broadband problems has been challenged by the prohibitively expensive computational cost of training neural networks. To break this barrier, this paper proposes a massively parallel strategy via frequency stacking. This architecture factorizes the broadband problem into independent, per-frequency branches that are trained in parallel, leveraging efficient GPU batching and parallel execution. Our results show that the proposed stacking architecture outperforms the conventional frequency conditioning architecture, with accelerations greater than two orders of magnitude and lower regression errors, which demonstrates a real-time potential for broadband sound field regression in small spaces. It is also demonstrated that incorporating a partial differential equation (PDE) residual loss further enhances the accuracy, particularly in the low to mid-frequency regime, despite at a cost of slightly longer training times. These findings offer practical guidelines for efficient, potentially real-time, yet high-fidelity sound field regression.