PathRIR: Physics-Guided Acoustic Path Selection and Late-Tail Compensation for Fast Room Impulse Response Simulation (IWAENC 2026)

Official implementation and pretrained models for our IWAENC 2026 paper.

PathRIR accelerates high-order image source method (ISM) simulation by pruning acoustically unimportant reflection paths. A lightweight Compensation-MLP restores the missing late-reverberation energy.

If you use this repository in your research, please cite:

@inproceedings{xu2026pathrir,
  author    = {Xu, Shaoheng and Sun, Chunyi and Zhang, Jihui and
               Bastine, Amy and Samarasinghe, Prasanga N. and
               Abhayapala, Thushara D.},
  title     = {{PathRIR}: Physics-Guided Acoustic Path Selection and Late-Tail
               Compensation for Fast Room Impulse Response Simulation},
  booktitle = {Proceedings of the 19th International Workshop on Acoustic
               Signal Enhancement (IWAENC)},
  year      = {2026}
}

• arXiv paper
• Personal academic website

• If you only want to try PathRIR to generate RIRs, install the package and use the example below.
• The remaining scripts are for rebuilding the datasets, training the models, and reproducing the evaluation.
• The bundled checkpoints use a maximum reflection order of 10, an 8 kHz sampling rate, and 0.5-second RIRs.
• Generated datasets are not included because the .npz files are large, but they can be recreated using the provided scripts.

PathRIR requires Python 3.9 or later.

git clone https://github.com/ShaoHenry/PathRIR.git
cd PathRIR

For pretrained inference:

python -m pip install .

For dataset generation, training, evaluation, and WAV export:

python -m pip install ".[full]"

The evaluation scripts were tested with pyroomacoustics==0.7.7. A GPU is optional. If you need a CUDA-enabled build of PyTorch, install it before installing PathRIR.

from pathrir import PathRIR

simulator = PathRIR()

rir = simulator.simulate(
    corners=[
        [0.0, 0.0],
        [5.0, 0.0],
        [6.0, 3.0],
        [3.0, 5.0],
        [0.0, 4.0],
    ],
    height=3.0,
    absorption=0.3,
    source=[2.0, 2.0, 1.5],
    mics=[
        [3.0, 3.0, 1.2],
        [1.5, 3.2, 1.6],
    ],
    fs=8000,
    duration=0.5,
    max_order=10,
)

print(rir.shape)  # (2, 4000)

Geometry and positions are measured in metres. corners defines the 2-D floor plan and may be listed clockwise or counter-clockwise.

absorption accepts:

• one value for all surfaces;
• one value per wall, with the wall mean used for the floor and ceiling; or
• one value per wall followed by separate floor and ceiling values.

Set compensate=False to return the pruning-only RIR. Set return_pruned=True to return both (compensated_rir, pruned_rir).

Custom checkpoints can be loaded with:

simulator = PathRIR(
    pruning_ckpt="path/to/pruning_checkpoint.pt",
    compensation_ckpt="path/to/compensation_checkpoint.pt",
)

New checkpoints are recommended when the room distribution, sampling rate, RIR duration, or maximum reflection order differs from the bundled settings.

See example_commands.txt for the complete command sequence. It covers:

1. dataset generation;
2. Pruning-MLP and Compensation-MLP training;
3. model evaluation; and
4. reflection-order evaluation from order 1 to 10.

High-order full-ISM simulation can require substantial memory and computation time. Test one room with one worker before starting a full run.

For timing comparisons, leave the machine otherwise idle, use at least three repeats, and add --no-mem-profiling.

Results are saved as per-room metrics, summary tables, and JSON files. Add --save-rir-wavs to export WAV files.

PathRIR is released under the MIT License. See LICENSE.