Abstract:

Few-shot room impulse response (RIR) prediction estimates target acoustics from sparse reference RIRs. Current models fuse references in a static Euclidean space, assuming a zero-curvature latent geometry for all temporal phases of the RIR. Using Gromov d-hyperbolicity, we show the acoustic manifold’s geometry evolves from a hierarchical tree structure during early reflections to a flat, diffuse late reverberation tail. Motivated by this heterogeneity, we propose Janus-RIR, a temporally gated, dual-geometry aggregation model. It employs a hyperbolic branch for early reflections and a Euclidean branch for smooth statistical averaging of the late tail, governed by a context-aware dynamic gate. Experiments on AcousticRooms show Janus-RIR achieves state-of-the-art. Aligning latent geometry with the physics of acoustics improves speech clarity (C50) while reducing late reverberation (T60) error by 26% over Euclidean baselines, adapting to environment-specific mixing times.