TEMPORAL AUDIO GROUNDING

Every sound.
In its own time.

Hearing what happens is only half the story. TAG-Bench asks when it happens—and whether you can find every occurrence.

Alibaba Token Foundry · TAG-Bench

ONE QUERY. EVERY OCCURRENCE.0534 / 1750

Eighteen metal clinks.

One minute of audio. Eighteen moments to find.

00:0000:3001:00
Annotated targets
A benchmark for locating sounds, words, meanings, and emotions in time.↓
1,750grounding queries
149.53 hof recorded audio
8listening tasks
22.1%multi-target queries
01 / THE BENCHMARK

Eight ways to ask when.

From a fraction of a second to twenty minutes. Explore the task mix, recording lengths, and how much of each recording contains the target.

Explore the dataset

The shape of temporal grounding

Recording length

Number of queries

Choose a duration range to filter all dataset charts.

How many moments?

Number of queries

Counts follow the original annotated intervals, before merging.

Coverage = annotated target time ÷ recording duration, using the union of intervals within the recording.

Browse the underlying data
IDTaskDurationTargetsTarget coverageAudio example
02 / THE CONTEXT CHALLENGE

Same sound.
A longer search.

These recordings contain the exact same 60-second audio segment. The target stays the same. The surrounding context grows.

THE SAME QUERY

Each timeline spans its full recording. The shaded region marks the embedded 60-second segment.

03 / THE LISTENING ROOM

Find it. Hear it. See it.

Sixteen real examples, two per task. Listen to complete recordings, jump to annotated moments, and explore the waveform or spectrogram.

ORIGINAL QUERY · ENGLISH

Annotated targetsClick the waveform to seek.

Annotated moments

Select an interval to play just that target. Timestamps use the original recording’s time origin.

Inspect the original query & annotation
FROM LISTENING TO LOCALIZATION

What. When. Every time.

TAG-Bench evaluates temporal audio grounding: given a recording and a natural-language query, return all matching time intervals. Accurate boundaries and complete occurrence recovery are both essential.

Read the paper ↗
CITATION

TAG-Bench: Benchmarking Temporal Audio Grounding in Large Audio Language Models

Yuhang Dai · Xin Shu · Zengxi Li · Lei Xie · Xiangang Li · Jianwei Yu
arXiv:2609.01542