README.md

metadata

pretty_name: ReasonAudio Natural
language:
  - en
license: cc-by-4.0
size_categories:
  - 1K<n<10K
tags:
  - audio-retrieval
  - multimodal-retrieval
  - reasoning
  - audioset
configs:
  - config_name: queries
    data_files:
      - split: test
        path: data/queries.csv
  - config_name: corpus
    data_files:
      - split: test
        path: data/corpus.csv
  - config_name: qrels
    data_files:
      - split: test
        path: data/qrels.csv
  - config_name: corpus_metadata
    data_files:
      - split: test
        path: data/corpus_metadata.csv

ReasonAudio Natural

ReasonAudio Natural is the real-world subtask of ReasonAudio, a benchmark for evaluating reasoning beyond semantic matching in text-audio retrieval. It contains 100 human-written queries and a retrieval corpus of 1,000 ten-second clips sampled from AudioSet-Strong. Each query may have multiple relevant clips.

Dataset statistics

Item	Count
Queries	100
Corpus entries	1,000
Positive relevance judgments	355
Mean positives per query	3.55
Median positives per query	1.0

Data files

• data/queries.csv: query IDs, query text, and subtask type.
• data/corpus.csv: corpus IDs and the filenames used by the ReasonAudio evaluation code.
• data/qrels.csv: multi-positive binary relevance judgments in query-id,corpus-id,score format.
• data/corpus_metadata.csv: AudioSet split, YouTube video ID, segment timestamps, and official AudioSet-Strong event annotations for every corpus entry.
• data/statistics.json: summary statistics for the released files.

The released qrels.csv is the multi-positive version used for Natural evaluation. It replaces the earlier one-anchor-per-query relevance file.

Loading

from datasets import load_dataset

queries = load_dataset("ReasonAudio/Natural", "queries", split="test")
corpus = load_dataset("ReasonAudio/Natural", "corpus", split="test")
qrels = load_dataset("ReasonAudio/Natural", "qrels", split="test")
metadata = load_dataset("ReasonAudio/Natural", "corpus_metadata", split="test")

Audio access

Raw audio is not redistributed in this repository. AudioSet is built from YouTube segments, and its official release provides identifiers, timestamps, labels, and derived features rather than the underlying source audio. The video_id, start_seconds, and end_seconds fields are provided so that researchers can identify the corresponding AudioSet-Strong segments subject to the original source terms and availability.

AudioSet annotations are provided by Google under CC BY 4.0, and the AudioSet ontology is provided under CC BY-SA 4.0. Copyright in the underlying source media remains with the respective rights holders.

• AudioSet: https://research.google.com/audioset/
• AudioSet-Strong: https://research.google.com/audioset/download_strong.html

The ReasonAudio-authored queries and relevance annotations are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

Annotation note

Natural queries combine sound-event identity with logical and temporal constraints, including negation, order, overlap, and duration. Relevance judgments were derived from AudioSet-Strong event metadata and audited during benchmark development. See the ReasonAudio paper for the complete construction and validation protocol.

Citation

@article{zhang2026reasonaudio,
  title={ReasonAudio: A Benchmark for Evaluating Reasoning Beyond Matching in Text-Audio Retrieval},
  author={Zhang, Honglei and Chen, Yuting and Hu, Chenpeng and Zhou, Pengfei and Zhang, Siyue and Shi, Yilei},
  journal={arXiv preprint arXiv:2605.03361},
  year={2026}
}