AudioSpan: Spanning the Duration and Depth of Audio Comprehension

Introduction

AudioSpan is a benchmark for long-form audio comprehension, spanning diverse durations and cognitive depths.

Questions come from two complementary paths:

• Native QA: questions drawn from the audio's natural content.
• Anchor QA: questions built around acoustic anchors planted into the audio.

Each path is scored in its own mode:

• Accuracy: multiple choice questions on native audio, scored by exact match.
• Rubric: open-ended questions on native audio, graded by rubric-based LLM judges against criteria.
• Chain: multiple-choice question chains on anchor audio; an answer is credited only up to the first error in the chain.

Quick Start

Download the release (~18 GB) and work from its root:

pip install -U huggingface_hub
hf download holvan/AudioSpan --repo-type dataset --local-dir AudioSpan
cd AudioSpan

1. Prepare Audio Data

Native recordings and sound events ship as tar archives (audio/audio_native.part*.tar, audio/audio_events.tar; the native parts are concatenated automatically). From the release root (the directory with prepare/ and audio/), verify them against audio/CHECKSUMS.sha256 and unpack into audio/:

# accuracy / rubric: native recordings only
python prepare/unpack_audio.py --buckets native
# -> audio/native/*.flac

# chain: also rebuild the anchor audio from the native recordings and the
# exact edit recipes in metadata/media/anchor/anchor_manifest.jsonl
# (needs only ffmpeg)
python prepare/unpack_audio.py --buckets native,events
python prepare/prepare_anchor.py
# -> audio/anchor/*.flac

(unpack_audio.py resolves the release root from its own location; pass --root if you run it from elsewhere.)

2. Prepare Inference Results

The questions live in metadata/{accuracy,rubric,chain}/{S,M,L}.jsonl; each record carries the question, its audio_path, and (for multiple choice) the four options. Run your model over each record and write one answer per question — only qa_id and the model's output; ground truth stays in the metadata and is joined in by the scorer:

accuracy / chain (multiple choice):

{"qa_id": "S_EN_001_P", "answer": "D"}

rubric (open-ended):

{"qa_id": "S_EN_001_P", "answer": "The narrator first says HDR at about 06:04 ..."}

• answer — the model's raw output; for multiple choice the scorer extracts the option letter (A/B/C/D) from it.
• Questions missing from your file count as wrong (or score zero).

3. Run Evaluation

# accuracy + chain: stdlib-only, no dependencies
python evaluate/score.py --mode accuracy --input your_accuracy.jsonl
python evaluate/score.py --mode chain    --input your_chain.jsonl

# rubric: needs an LLM judge (default: gpt-5.4-2026-03-05; API key from the
# matching provider env var, e.g. OPENAI_API_KEY; or pass --api-base/--api-key)
pip install openai tenacity
export OPENAI_API_KEY="your-api-key-here"
python evaluate/score_rubric.py --input your_rubric.jsonl

Both scorers join your answers against the ground truth in metadata/ by qa_id — no reference answers needed in your submission.

License and Data Use

AudioSpan is released for non-commercial research and evaluation. Our artifacts (QA items, rubric criteria, anchor manifests, sound events, and evaluation scripts) are under CC BY-NC-SA 4.0; copyright of the source recordings remains with their original creators. The recordings are included only for evaluation, and downloading the dataset constitutes agreement not to redistribute the audio or use it in commercial products.

Citation

If you find AudioSpan useful for your research, please consider citing:

@article{huang2026audiospan,
  title   = {AudioSpan: Spanning the Duration and Depth of Audio Comprehension},
  author  = {Huang, Wen and Chu, Yunfei and Gao, Meng and He, Haolin and Xu, Jin},
  journal = {arXiv preprint arXiv:2608.26431},
  year    = {2026}
}