File size: 1,802 Bytes
ce0dfcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d54cc87
ce0dfcc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: mit
language:
- en
tags:
- audio
- temporal-grounding
- audio-language-modeling
pretty_name: SpotSound-Bench
size_categories:
- n<1K
---

# SpotSound-Bench: A 'Needle-in-a-Haystack' Evaluation for Audio Temporal Grounding

[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://loiesun.github.io/spotsound/)
[![GitHub](https://img.shields.io/badge/GitHub-Code-black)](https://github.com/LoieSun/spotsound)
[![Paper](https://img.shields.io/badge/arXiv-Paper-red)](https://arxiv.org/abs/2604.13023)

## Benchmark Summary

**SpotSound-Bench** is a challenging temporal grounding benchmark designed to evaluate Large Audio-Language Models (ALMs). 

Existing benchmarks for audio temporal grounding often feature high ratios of target-window duration to full audio clip duration, which fail to simulate real-world scenarios where short events are obscured by dense background sounds. To bridge this gap, we introduce SpotSound-Bench, featuring short acoustic events embedded within long, unstructured recordings. 

This benchmark creates a rigorous **‘needle-in-a-haystack’** evaluation, demanding high temporal precision and robust resistance against hallucinations from audio-language models.

## Benchmark Characteristics

- **Average Clip Length:** 53.4 seconds
- **Average Target Event Length:** 4.5 seconds
- **Temporal Density:** 8.4% (Target event duration / Full audio clip duration)
- **Challenge:** A large search space dominated by background content, requiring models to pinpoint exact timestamps of short events while ignoring complex background ambiance and avoiding hallucinated predictions for non-existent events.

## Data Structure

<pre>
{
    "audio_path": "audio/-1q1otOq9TU_315_345.wav",
    "caption": "skidding",
    "gt": [[11.2, 20.3]]
}
</pre>