AcousticAgent is a simulation platform that starts from structured floorplans and turns them into editable, acoustically connected multi-room scenes. The authors note that existing tools either generate room impulse responses quickly in parameterized rooms or rely on preconstructed 3D scenes, making it difficult to build many editable residential environments. The new system jointly models room geometry and connectivity, furniture, materials, sound sources, and microphone configurations, and uses a floorplan-aware hybrid geometrical-acoustics solver to produce multichannel room impulse responses and reverberant audio.

To support downstream research, the authors built FP-RIR, a dataset of 15,376 residential floorplans covering varied room topologies, propagation conditions, and source-receiver configurations. The platform also supports directional and moving sources, distributed microphones, and vision-language-model-assisted scene construction from floorplan images or text descriptions.

The paper validates the simulated acoustics against real-world measurements and evaluates whole-home source localization, speech enhancement, and far-field speech processing in both single-device and distributed microphone settings. The results indicate that floorplan-aware modeling improves fidelity and that distributed spatial processing helps in multi-room environments. The authors position AcousticAgent and FP-RIR as scalable, configurable resources for developing and evaluating residential speech systems.