File size: 3,178 Bytes
9c49f91
 
 
 
 
b8b7f57
9c49f91
f8e26d6
2fe5b3c
 
 
 
 
 
 
 
 
75a11d7
2fe5b3c
 
 
f3769c7
 
 
b850911
 
 
2472210
b850911
ceb0e43
b850911
 
ceb0e43
2472210
ceb0e43
3a0df9b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
386fa87
3a0df9b
 
 
386fa87
3a0df9b
 
 
 
 
3fe0910
 
 
 
 
 
 
8be54bf
 
 
 
 
 
 
 
 
 
 
 
 
3fe0910
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
---
license: cc-by-nc-4.0
language:
- zh
- en
viewer: false
---
[![arXiv](https://img.shields.io/badge/arXiv-2604.22245-b31b1b.svg)](https://arxiv.org/abs/2604.22245)
# LAT-Bench
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:
- **Dense Audio Captioning (DAC)**: generate temporally grounded descriptions over the full audio
- **Temporal Audio Grounding (TAG)**: localize relevant time spans for a given query
- **Targeted Audio Captioning (TAC)**: produce descriptions for specific temporal segments

LAT-Bench contains approximately **40 hours of long-form audio**, including:
- **25 hours in Chinese**
- **15 hours in English**

The dataset spans diverse real-world scenarios, including conversations, lifestyle vlogs, educational content, and so on.

## Data Distribution

<p align="center">
  <img src="./Figures/bench-figure.png" width="600"/>
  <em>Figure 1: Duration and scenario distributions of LAT-Bench across Chinese and English.</em>
</p>
LAT-Bench exhibits balanced coverage across duration ranges and scenarios, ensuring robust evaluation under diverse long-form settings.

<p align="center">
  <em>Table 1: Temporal annotation statistics of LAT-Bench across DAC, TAG, and TAC tasks.</em>
  <img src="./Figures/bench-table.png" width="600"/>
</p>
The annotations provide comprehensive temporal coverage across the beginning, middle, and end of audio sequences.

## Data Organization

LAT-Bench is organized into two types of files: **metadata files** and **task files**.

### Metadata Files

- `./meta/bench-CN-meta.jsonl`
- `./meta/bench-EN-meta.jsonl`

These files provide metadata for each audio sample, including:
- `id`: unique identifier
- `url`: source link for downloading the audio
- `title`: original audio title
- `duration`: duration in seconds

### Task Files
Dense Audio Captioning (DAC)
- `./task/bench-CN-DAC.jsonl`
- `./task/bench-EN-DAC.jsonl`
  
Temporal Audio Grounding (TAG)
- `./task/bench-CN-TAG.jsonl`
- `./task/bench-EN-TAG.jsonl`
  
Targeted Audio Captioning (TAC)
- `./task/bench-CN-TAC.jsonl`
- `./task/bench-EN-TAC.jsonl`

Each task file contains benchmark instances in a unified format.
The audios field references the corresponding audio sample using the id from metadata files.

## Evaluation Protocol

For detailed evaluation protocols and metrics, please refer to the official repository:  
👉 https://github.com/alanshaoTT/LAT-Audio-Repo

## Citation

If you find this work useful, please cite:

```bibtex
@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}
}
```

## Contact

For questions, feedback, or collaboration inquiries, please contact:

📧 mcshao@mail.nwpu.edu.cn