Long-HEAR

Status: coming soon. The Long-HEAR data is currently going through our release review process. Files will be uploaded to this repository as soon as the review is complete. Follow this repository to be notified when the data becomes available.

Long-HEAR is the dataset and benchmark accompanying the paper ChronoAudio: Learning Temporal Audio Reasoning from Sound Interventions. It targets temporal audio question answering, where both the answer and its supporting evidence (the relevant sound events with their onsets and offsets) must follow the timing of sounds in the recording.

Planned contents

The release is planned to include:

• Audio scenes with complete sound-event annotations (event type, onset, offset).
• Temporal question-answer pairs with complete evidence annotations, including long-form inputs.
• Source-relocation edits for testing whether answers and evidence respond correctly when individual sound sources are moved in time.
• Documentation of the data construction process.

Components may be released in stages; each component will be marked here when it becomes available. Exact statistics, split definitions, file formats, and loading instructions will be documented together with the data.

Source corpora

Long-HEAR is built from publicly available corpora: FSD50K, LibriSpeech, and MUSAN for synthetic mixtures, and AudioSet-Strong, DESED, and the AMI Meeting Corpus for natural recordings.

License

The license will be announced with the release. Because Long-HEAR is derived from third-party corpora, the released data will follow the licenses and attribution requirements of its upstream sources, some of which (e.g., part of FSD50K) restrict commercial use.

Citation

@misc{hu2026chronoaudio,
  title  = {ChronoAudio: Learning Temporal Audio Reasoning from Sound Interventions},
  author = {Hu, Zhiyuan and Zhang, Qiyu and Yang, Xusheng and Zou, Yuexian},
  year   = {2026}
}

Contact

Please post questions in the Community tab of this repository.