Abstract:

Room impulse responses (RIRs) characterize sound propagation in a scene, determined by its geometry and materials. Traditional RIR estimation relies on acoustic simulation, while recent deep learning focuses on spatial interpolation. However, acoustic simulation requires material annotation of each object, and interpolation requires multiple RIRs whenever the scene is edited slightly (e.g., rearranging furniture). Both approaches are inefficient and impractical in everyday environments. In this paper, we introduce edit-conditioned RIR estimation, which updates a measured RIR to reflect geometry edits. To relate the geometry edits to acoustic variation without material annotations, we simulate proxy RIRs from the room meshes before and after the edit using predefined materials. We propose PG-RIR, which leverages these proxies to estimate the variation in the real RIR. Experiments show that PG-RIR accurately predicts RIRs under furniture rearrangement and removal.