From CVPR 2025 Paper: Hearing Anywhere in Any Environment
Authors:
Xiulong Liu¹, Anurag Kumar², Paul Calamia², Sebastia V. Amengual², Calvin Murdock²,
Ishwarya Ananthabhotla², Philip Robinson², Eli Shlizerman¹, Vamsi Krishna Ithapu², Ruohan Gao³
¹University of Washington, ²Meta, ³University of Maryland, College Park
AcousticRooms is a large-scale synthetic room impulse response (RIR) dataset designed for cross-room RIR prediction tasks. It includes over 300,000 single-channel RIRs simulated across 260 rooms spanning 10 categories, such as apartment, auditorium, office, and cafe. Each room features high-quality 3D spatial geometry and randomized material properties drawn from a diverse library of 332 acoustic materials across 11 categories.
- 260 professionally designed 3D room geometries from 10 room types
- Over 300K high-quality RIRs simulated using hybrid acoustic simulation (wave + geometric)
- 332 real-world acoustic materials randomly assigned to room surfaces
- Diverse room sizes (20 m³ – 1000 m³) and layouts
- Source-receiver configurations with spatially valid random sampling
Simulations are run using the Treble acoustic simulation platform in hybrid mode:
- Wave-based simulation for low frequencies (< 710 Hz)
- Geometric-based simulation for high frequencies (> 710 Hz), using:
- Image-source method up to order 4 with 50k rays
- Stochastic ray-tracing for higher orders with 5000 rays
- RIRs simulated until 60dB energy decay
- 10–100 omnidirectional sources and 25–100 monaural receivers per room
- Placement constraints:
- ≥ 0.5m from surfaces
- ≥ 1.0m between sources
- ≥ 0.5m between sources and receivers
- Valid locations sampled using spatial point-picking algorithm
- Surfaces (walls, floor, ceiling, furniture) are semantically labeled
- Each label mapped to one of 11 material categories
- Randomized selection of specific material with realistic acoustic coefficients from a library of 332 materials
- Diverse acoustic properties across geometrically similar rooms
This repository includes the following files and folders:
| File/Folder | Description |
|---|---|
single_channel_ir.zip |
300K normalized, single-channel RIR waveforms (with 22.05kHz sampling rate and trimed) across all 10 room categories. Naming: S{sourceID}_R{receiverID}_hybrid_IR.wav. |
metadata.zip |
JSON metadata for each RIR pair, including [x, y, z] positions of source and receiver |
depth_map.zip |
Panoramic (equirectangular) depth maps captured at each receiver location, each depth map is named as {receiverID}.npy |
room_mesh_obj_format/ |
3D mesh geometries of all rooms in .obj format |
material_library/ |
Acoustic material definitions including absorption/reflection coefficients |
simulation_info/ |
Configuration files and simulation parameters for selected rooms |
raw_dense_simulation/ |
A subset of original simulated single-channel RIRs without any processing (sampled at 32kHz) along with corresponding room geometry (room_id different from preprocessed RIRs). |
AcousticRooms is used as a benchmark for cross-room RIR prediction task in the paper. If you are interested in this task, please check out the third-party implementation of the xRIR framework proposed in the CVPR paper at https://github.com/DragonLiu1995/xRIR_code.
AcousticRooms is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
You are free to share and adapt the dataset, even for commercial use, as long as proper attribution is given.
See the LICENSE file for full terms.
If you use this dataset in your research, please cite the following paper:
@inproceedings{liu2025haae,
title = {Hearing Anywhere in Any Environment},
author = {Liu, Xiulong and Kumar, Anurag and Calamia, Paul and Garí, Sebastià V. Amengual and Murdock, Calvin and Ananthabhotla, Ishwarya and Robinson, Philip and Shlizerman, Eli and Ithapu, Vamsi Krishna and Gao, Ruohan},
booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2025},
}