LAT-Bench is the first benchmark designed for evaluating temporal awareness in long-form audio understanding. Unlike existing benchmarks limited to short clips, LAT-Bench supports audio durations up to 30 minutes, enabling evaluation under realistic long-form scenarios.
The benchmark covers three core tasks:
LAT-Bench contains approximately 40 hours of long-form audio, including:
The dataset spans diverse real-world scenarios, including conversations, lifestyle vlogs, educational content, and so on.
Figure 1: Duration and scenario distributions of LAT-Bench across Chinese and English.
Table 1: Temporal annotation statistics of LAT-Bench across DAC, TAG, and TAC tasks.
LAT-Bench is organized into two types of files: metadata files and task files.
./meta/bench-CN-meta.jsonl./meta/bench-EN-meta.jsonlThese files provide metadata for each audio sample, including:
id: unique identifierurl: source link for downloading the audiotitle: original audio titleduration: duration in secondsDense Audio Captioning (DAC)
./task/bench-CN-DAC.jsonl./task/bench-EN-DAC.jsonlTemporal Audio Grounding (TAG)
./task/bench-CN-TAG.jsonl./task/bench-EN-TAG.jsonlTargeted Audio Captioning (TAC)
./task/bench-CN-TAC.jsonl./task/bench-EN-TAC.jsonlEach task file contains benchmark instances in a unified format. The audios field references the corresponding audio sample using the id from metadata files.
For detailed evaluation protocols and metrics, please refer to the official repository:
๐ https://github.com/alanshaoTT/LAT-Audio-Repo
If you find this work useful, please cite:
@article{shao2026lataudio,
title={Listening with Time: Precise Temporal Awareness for Long-Form Audio Understanding},
author={Shao, Mingchen and Su, Hang and Tian, Wenjie and Mu, Bingshen and Lin, Zhennan and Fan, Lichun and Luo, Zhenbo and Luan, Jian and Xie, Lei},
journal={arXiv preprint arXiv:2604.22245},
year={2026}
}
For questions, feedback, or collaboration inquiries, please contact:
26 commits
LAT-Bench is the first benchmark designed for evaluating temporal awareness in long-form audio understanding. Unlike existing benchmarks limited to short clips, LAT-Bench supports audio durations up to 30 minutes, enabling evaluation under realistic long-form scenarios.
The benchmark covers three core tasks:
LAT-Bench contains approximately 40 hours of long-form audio, including:
The dataset spans diverse real-world scenarios, including conversations, lifestyle vlogs, educational content, and so on.
Figure 1: Duration and scenario distributions of LAT-Bench across Chinese and English.
Table 1: Temporal annotation statistics of LAT-Bench across DAC, TAG, and TAC tasks.
LAT-Bench is organized into two types of files: metadata files and task files.
./meta/bench-CN-meta.jsonl./meta/bench-EN-meta.jsonlThese files provide metadata for each audio sample, including:
id: unique identifierurl: source link for downloading the audiotitle: original audio titleduration: duration in secondsDense Audio Captioning (DAC)
./task/bench-CN-DAC.jsonl./task/bench-EN-DAC.jsonlTemporal Audio Grounding (TAG)
./task/bench-CN-TAG.jsonl./task/bench-EN-TAG.jsonlTargeted Audio Captioning (TAC)
./task/bench-CN-TAC.jsonl./task/bench-EN-TAC.jsonlEach task file contains benchmark instances in a unified format. The audios field references the corresponding audio sample using the id from metadata files.
For detailed evaluation protocols and metrics, please refer to the official repository:
๐ https://github.com/alanshaoTT/LAT-Audio-Repo
If you find this work useful, please cite:
@article{shao2026lataudio,
title={Listening with Time: Precise Temporal Awareness for Long-Form Audio Understanding},
author={Shao, Mingchen and Su, Hang and Tian, Wenjie and Mu, Bingshen and Lin, Zhennan and Fan, Lichun and Luo, Zhenbo and Luan, Jian and Xie, Lei},
journal={arXiv preprint arXiv:2604.22245},
year={2026}
}
For questions, feedback, or collaboration inquiries, please contact:
26 commits