TasnimKabir12/audita-audio

AUDITA: Audio Question Answering Benchmark

AUDITA is a benchmark of human-authored audio question-answer pairs designed to evaluate auditory reasoning in real-world settings. Unlike synthetic or template-based audio QA datasets, AUDITA focuses on human-written trivia-style questions that require grounding in acoustic, temporal, and semantic cues.

The dataset is designed to expose limitations of current audio-language models, particularly their reliance on linguistic priors and weak audio grounding.

📄 Paper

AUDITA: A New Dataset to Audit Humans vs. AI Skill at Audio QA

ACL Findings 2026: https://aclanthology.org/2026.findings-acl.1292/

arXiv: https://arxiv.org/abs/2604.21766

🔎 Dataset Explorer

An interactive version of the dataset is available at:

https://manchester.umiacs.umd.edu/audio

The explorer allows you to:

• Browse examples by source dataset
• Browse examples by audio category
• Listen to audio clips
• View question-answer pairs and metadata

---

📁 Dataset Overview

AUDITA consists of two dataset splits.

sources/ — Human-Authored Audio Questions

This is the primary contribution of AUDITA.

It contains 6,460 human-authored audio question-answer pairs collected from real-world trivia and competitive quiz sources, including:

• Quizmasters
• PAVEMENT
• Audio-Packets

Key properties

• Human-written questions
• Real-world audio grounding
• Requires auditory reasoning beyond simple sound recognition
• Covers music, speech, media, and environmental sounds
• Includes both open-ended and closed-ended questions

---

external/ — Existing Audio QA Benchmarks

This split contains 3,230 examples from existing audio QA datasets:

• OpenAQA
• ClothoAQA

These datasets are included to:

• Provide comparison with prior work
• Evaluate generalization across benchmark styles
• Highlight differences in question difficulty and reasoning requirements

They are included as evaluation baselines rather than the primary contribution of AUDITA.

---

📊 Dataset Statistics

Split	Examples
Human-authored (sources)	6,460
External (external)	3,230
Total	9,690

The dataset spans multiple audio domains, including music, speech, media content, environmental sounds, and sound identification.

---

🔊 Data Format

Each example has the following structure:

{
    "audio": Audio(),
    "question": str,
    "answer": str,
    "source": str,
    "category": str
}

Field	Description
audio	Audio clip
question	Natural-language question about the audio
answer	Ground-truth answer
source	Dataset split (sources or external)
category	High-level semantic category

---

📖 Citation

If you use AUDITA in your research, please cite the ACL Findings paper:

@inproceedings{kabir-etal-2026-audita,
    title = "{AUDITA}: A New Dataset to Audit Humans vs. {AI} Skill at Audio {QA}",
    author = "Kabir, Tasnim and
      Kurdydyk, Dmytro and
      Palnitkar, Aadi and
      Dorn, Liam and
      Ahmed, Ahmed Haj and
      Boyd-Graber, Jordan Lee",
    editor = "Liakata, Maria and
      Moreira, Viviane P. and
      Zhang, Jiajun and
      Jurgens, David",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2026",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.findings-acl.1292/",
    pages = "25922--25951",
    ISBN = "979-8-89176-395-1"
}

Alternatively, you may cite the arXiv preprint:

@article{kabir2026audita,
  title={AUDITA: A New Dataset to Audit Humans vs. AI Skill at Audio QA},
  author={Kabir, Tasnim and Kurdydyk, Dmytro and Palnitkar, Aadi and Dorn, Liam and Ahmed, Ahmed Haj and Boyd-Graber, Jordan Lee},
  journal={arXiv preprint arXiv:2604.21766},
  year={2026}
}