README.md

metadata

license: cc-by-4.0
task_categories:
  - audio-classification
tags:
  - spatial-audio
  - room-impulse-response
  - audio-representation-learning
  - probing
pretty_name: SARL Spatial Audio RIRs

SARL — Spatial Audio RIRs

Room impulse responses for the SARL benchmark (Probing Spatial Structure in Pretrained Audio Representations, accepted at Interspeech 2026; arXiv:2606.05544). Code and full instructions: https://github.com/chuyangchencd/SARL

This holds only the RIRs. The source clips (audio/) and ambient noise (ambient/) can't be redistributed, so you build them from public datasets following the code repo.

Contents

The RIRs ship as 18 tar archives — one per (family × split × format) — plus a small metadata.json per split (not tarred, so it's directly browsable):

rir_source_<split>_<fmt>.tar     source-task RIRs (ray-traced) -> rir_source/<split>/<fmt>/NNNNN.npy
rir_room_<split>_<fmt>.tar       room-task RIRs (shoebox)      -> rir_room/<split>/<fmt>/NNNNN.wav
rir_source/<split>/metadata.json   one per split, shared by all formats
rir_room/<split>/metadata.json     one per split, shared by all formats

• <split> ∈ train / val / test (rooms disjoint across splits).
• <fmt> is the stored capture format: mic (4-channel tetrahedral; downmixed to stereo on load), binaural (2-channel), foa (4-channel first-order Ambisonics, ACN/SN3D — AmbiX). Every scene is provided in all three.
• rir_source — .npy of shape [C, N, T] (N candidate source positions); metadata gives per-position coordinates (azimuth −180–180°, elevation −60–60°, distance 0.5–2.5 m) and space_name.
• rir_room — .wav of shape [C, T]; metadata gives rt60 (s), volume (m³), and shape (0–3 = cube/flat/corridor/rectangular).

Usage

Download and unpack the tars in place:

huggingface-cli download chuyangchenn/SARL --repo-type dataset --local-dir data_root
cd data_root && for t in rir_*.tar; do tar xf "$t"; done && rm rir_*.tar

To fetch only the capture format you need (e.g. foa), add --include "*_foa.tar" "*/metadata.json" "README.md" to the download.

Then build audio/ and ambient/ into the same data_root (see the code repo) and run the benchmark. These RIRs are the exact ones used in the paper.

Citation

@article{chen2026sarl,
  title   = {Probing Spatial Structure in Pretrained Audio Representations},
  author  = {Chen, Chuyang and Ding, Sivan and Roman, Adrian S. and Bello, Juan P.},
  journal = {arXiv preprint arXiv:2606.05544},
  year    = {2026},
}