Velocity Potential Neural Field for Efficient Ambisonics Impulse Response Modeling

A team from Mitsubishi Electric Research Laboratories (MERL) has introduced a novel AI architecture for modeling 3D spatial audio, accepted for presentation at ICASSP 2026. Their 'Velocity Potential Neural Field' tackles the challenge of reconstructing First-Order Ambisonics (FOA) signals, a standard format for immersive audio. Unlike previous physics-informed neural networks that used soft penalty terms to encourage physical plausibility, this model is designed from the ground up to be physically correct. It does this by having the neural network approximate a single, foundational scalar function called the velocity potential, rather than the four-channel audio signal itself.

The key innovation is that the four components of the FOA signal—which represent sound pressure and particle velocity in 3D space—are mathematically derived from the partial derivatives of this learned potential. This means the final audio output inherently satisfies the linearized momentum equation, a core principle of acoustics, at any point in time and space. The researchers validated their framework on the task of room impulse response reconstruction, a critical component for realistic audio simulation and virtual reality. By enforcing hard physical constraints through the model's architecture, they achieve more efficient and accurate spatial audio interpolation from sparse microphone measurements, paving the way for higher-fidelity immersive sound experiences in AR/VR and acoustic design.