Scene-adaptive audio in extended reality (XR) needs acoustic parameters that match the user's surroundings, but running full geometric-acoustic simulations for every scene update is computationally expensive. The authors of Point2Acoustic propose a learning-based alternative: represent a room as a surface-sampled point cloud, where each point carries spatial coordinates plus frequency-dependent absorption and scattering coefficients, then feed that representation into a PointNet or PointNet++ encoder to directly output acoustic parameters.
The model was trained and evaluated on a synthetic furnished-room dataset with varied geometry and materials, using geometric-acoustic simulation to generate reference targets. It predicts six-band reverberation time (RT60) and reverberant-to-direct ratio (RDR) accurately, with inference at millisecond scale. The authors also ran ablations to test whether adding explicit global room descriptors or semantic surface labels helped; neither provided a consistent benefit, suggesting the point-cloud representation alone captures the relevant information.
The work is submitted to IEEE ICASSP 2027 and is currently available on arXiv. If the approach generalizes beyond synthetic scenes, it could make runtime acoustic adaptation practical for immersive audio, letting XR systems update sound rendering as users move through or modify their environment without repeated simulation costs.