By requesting access, you confirm that your use is limited to non-commercial academic purposes and that you have read and agree to the Matterport3D Terms of Use: https://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf

ST-AudioQA

ST-AudioQA is the benchmark for Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources.

Request access on this page and run hf auth login before downloading the dataset.

Split	Recordings	QA pairs
Train	40,000	734,402
Validation	4,000	73,465
Test	10,000	183,834

The dataset contains paper-aligned QA annotations, AudioSet segment references, complete acoustic recipes, and 1,958 deduplicated exact test RIRs. Original AudioSet waveforms and Matterport3D scene assets are not redistributed.

Load the tables

from datasets import load_dataset

qa = load_dataset("HBoh/ST-AudioQA", "qa")
recordings = load_dataset("HBoh/ST-AudioQA", "recordings")
sources = load_dataset("HBoh/ST-AudioQA", "sources")
rir_recipes = load_dataset("HBoh/ST-AudioQA", "rir_recipes")
rir_test_index = load_dataset("HBoh/ST-AudioQA", "rir_test_index")

qa contains questions, answers, QA types, families, time anchors, and target metadata.
recordings describes each virtual recording, its sources, scene, and listener position.
sources provides AudioSet/YouTube IDs, source segments, labels, gains, and RIR IDs.
rir_recipes provides scene geometry, trajectories, timing, and renderer parameters.
rir_test_index maps each exact test rir_id to a tar shard and member path.

Exact test RIRs

The complete repository is approximately 58 GB, mostly in 28 uncompressed test-RIR shards. Use the extraction helper from the ST-AudioLM repository to download only the shard containing a requested RIR:

python scripts/extract_test_rir.py 6fbce3a3a725a435ce1090e3 \
  --output rirs/6fbce3a3a725a435ce1090e3.npz

Use the exact test RIRs for benchmark reproduction. Train and validation RIRs can be regenerated from the released recipes and pinned renderer setup, but Habitat-Sim re-rendering is not guaranteed to be bit-identical across builds.

Reconstruction and RIR-rendering code is provided in the ST-AudioLM repository. Reconstructed model inputs are 10-second, 32-kHz AmbiX ACN/SN3D recordings in [W, Y, Z, X] order.

License and source assets

ST-AudioQA annotations and released RIR assets are provided under CC BY-NC-SA 3.0 US and remain subject to the Matterport3D Terms of Use where applicable. Users must obtain AudioSet source recordings and Matterport3D assets from their original providers and comply with all applicable upstream terms.

Limitations

Referenced AudioSet source videos can become unavailable or change over time.
Event labels inherit AudioSet ontology coverage and annotation noise.
Simulated Matterport3D environments do not represent every real acoustic condition.
Regenerated RIRs can vary with simulator builds; exact test RIRs are supplied for stable evaluation.

Citation

@article{hyun2026spatio,
  title={Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources},
  author={Hyun-Bin, Oh and Shimada, Kazuki and Takida, Yuhta and Sung-Bin, Kim and Uesaka, Toshimitsu and Shibuya, Takashi and Lee, Kyeongyoon and Oh, Tae-Hyun and Mitsufuji, Yuki},
  journal={arXiv preprint arXiv:2606.14141},
  year={2026}
}

Downloads last month: 11

Total file size:

58 GB

Paper for HBoh/ST-AudioQA

Spatio-Temporal Audio Language Modeling for Dynamic Sound Sources

Paper • 2606.14141 • Published Jun 12