saeedzou/rir-former-datasets

RIR-Former Datasets

Synthetic room impulse response (RIR) data for training and evaluating RIR-Former–style models. Each example pairs a simulated multichannel RIR with the room and array geometry that produced it, so a model can learn the mapping between acoustic geometry and the resulting impulse response.

The dataset is organized into two configs, exp1 and exp2, corresponding to two different simulation setups (see Configs below). Both configs share the same feature schema and split structure (train / val / test).

Dataset Summary

• Task: room acoustics / impulse response modeling, useful for RIR estimation, dereverberation, acoustic scene simulation, and geometry-conditioned audio generation.
• Source: simulated (synthetic), not recorded in real rooms.
• Splits: train, val, test for each config.
• Configs: exp1, exp2 — same fields, different simulation scale (exp2 examples are roughly 2x the size of exp1 examples on disk, e.g. more microphones and/or longer RIR length; update this line with the exact simulation parameters used for each experiment).

Supported Tasks

• RIR regression / generation: predict rir from room_dims, mic_positions, src_position, and/or rt60.
• RT60 estimation: predict rt60 from rir and geometry.
• Geometry-conditioned acoustic simulation: use room and array parameters to synthesize plausible RIRs.

Dataset Structure

Data Fields

Field	Type	Description
rir	list[list[float32]]	Simulated room impulse response(s). Outer list indexes channels/microphones, inner list holds the time-domain samples for that channel.
mic_positions	list[list[float32]]	3D coordinates [x, y, z] for each microphone in the array.
src_position	list[float32]	3D coordinate [x, y, z] of the sound source.
room_dims	list[float32]	Room dimensions, e.g. [length, width, height] in meters.
rt60	float32	Reverberation time (RT60) of the simulated room, in seconds.

Configs

exp1

Baseline simulation setup. Smaller per-example footprint (fewer microphones and/or shorter RIRs than exp2). Splits: train 8,000 / val 200 / test 10. Download size ~1.97 GB, dataset size ~2.16 GB.

exp2

Larger-scale simulation setup, roughly double the per-example data size of exp1 (e.g. a larger microphone array and/or longer RIR length — confirm and describe the exact difference from your simulation config). Splits: train 8,000 / val 200 / test 10. Download size ~4.32 GB, dataset size ~4.32 GB.

Note: fill in the precise simulation parameters that differ between exp1 and exp2 (number of microphones, array geometry, RIR length/sample rate, room size ranges, RT60 ranges, source/mic placement constraints, etc.) so users know exactly what each config represents.

Usage

from datasets import load_dataset

# Load the exp1 config
ds = load_dataset("saeedzou/rir-former-datasets", "exp1")

# Load the exp2 config
ds = load_dataset("saeedzou/rir-former-datasets", "exp2")

Each config exposes train, val, and test splits:

ds["train"][0]
# {
#   "rir": [[...], [...], ...],
#   "mic_positions": [[x1, y1, z1], [x2, y2, z2], ...],
#   "src_position": [x, y, z],
#   "room_dims": [length, width, height],
#   "rt60": 0.42,
# }

Dataset Creation

Room impulse responses were simulated using rir-generator, an image-source method implementation for generating RIRs. For each example, a room geometry, microphone array layout, and source position were sampled, and the corresponding RIR(s) and RT60 were computed from the simulation.

The exact generation script used to build both configs is available here: generate_full_datasets.py, part of the RIR-Former project.

Note: add the specific sampling ranges used for room dimensions, RT60, source/mic positions, and array geometry, if you'd like to document them here directly.

Considerations for Using the Data

• This is synthetic data; models trained only on these RIRs may not fully generalize to real-room recordings without additional real-world fine-tuning or domain adaptation.
• The val and test splits are intentionally small (200 and 10 examples per config); treat test-split metrics as indicative rather than statistically robust, and consider held-out real recordings for final evaluation if available.

Citation

If you use this dataset, please cite:

@inproceedings{Xu_2026,
   title={RIR-Former: Coordinate-Guided Transformer for Continuous Reconstruction of Room Impulse Responses},
   url={http://dx.doi.org/10.1109/ICASSP55912.2026.11462487},
   DOI={10.1109/icassp55912.2026.11462487},
   booktitle={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
   publisher={IEEE},
   author={Xu, Shaoheng and Sun, Chunyi and Zhang, Jihui Aimee and Samarasinghe, Prasanga and Abhayapala, Thushara},
   year={2026},
   month=May,
   pages={15312–15316}
}

Links

• Project repository: github.com/saeedzou/RIR-Former
• Author GitHub: github.com/saeedzou
• RIR simulator used: github.com/audiolabs/rir-generator