Physics-regularized neural acoustic fields for spatial layout inference from sparse room impulse responsesa)

Sparse room impulse responses (RIRs) encode indoor spatial structure through direct sound, early reflections, and reverberation shaped by room boundaries and objects. However, estimating a spatial layout from only a few source-receiver RIR measurements is highly ill-posed because the acoustic field is spatially undersampled and different layouts may produce similar echo patterns. This problem is important for acoustic sensing, robotic navigation, and environmental perception when dense measurements or reliable visual observations are unavailable. We propose a two-stage framework for spatial layout inference from sparse RIRs. First, an implicit neural acoustic field reconstructs a continuous acoustic representation from sparse source-receiver responses. Acoustic reciprocity is introduced as an unsupervised physical regularizer to improve interpolation in unobserved regions, and the reconstructed field is represented in the Mel-spectrogram domain as a compact intermediate feature. Second, a cross-modal acoustic-to-visual network translates this reconstructed field into an image-based spatial layout proxy. Experiments on high-fidelity simulations and a public indoor acoustic benchmark show improved acoustic field reconstruction and reliable spatial layout inference from sparse RIR measurements.