---
pretty_name: AudioVerse
language:
- en
tags:
- audio
- text
- annotations
- audio-text-retrieval
configs:
- config_name: all
  default: true
  data_files:
  - split: train
    path: "data/*/*.parquet"
- config_name: audioset
  data_files:
  - split: train
    path: "data/audioset/*.parquet"
- config_name: audioset_sl
  data_files:
  - split: train
    path: "data/audioset_sl/*.parquet"
- config_name: audiostock
  data_files:
  - split: train
    path: "data/audiostock/*.parquet"
- config_name: bbc_sound_effects
  data_files:
  - split: train
    path: "data/bbc_sound_effects/*.parquet"
- config_name: epic_sound
  data_files:
  - split: train
    path: "data/epic_sound/*.parquet"
- config_name: free_to_use_sounds
  data_files:
  - split: train
    path: "data/free_to_use_sounds/*.parquet"
- config_name: freesound
  data_files:
  - split: train
    path: "data/freesound/*.parquet"
- config_name: sonniss_game_effects
  data_files:
  - split: train
    path: "data/sonniss_game_effects/*.parquet"
- config_name: soundbible
  data_files:
  - split: train
    path: "data/soundbible/*.parquet"
- config_name: vggsound
  data_files:
  - split: train
    path: "data/vggsound/*.parquet"
---

# AudioVerse

AudioVerse provides **1.4M audio annotation records for audio-text retrieval**.
The dataset is described in [Re-purposing Multimodal Large Language Models for Audio-Text Retrieval](https://arxiv.org/abs/2602.18010).

This release contains text annotations only; audio files are not included.

## Data format

Annotations are stored in Parquet files with five string fields:

| Field | Description |
|---|---|
| `dataset` | Source dataset name. |
| `id` | Sample identifier within the source dataset. |
| `long_caption` | Detailed description of the audio content. |
| `short_caption` | Concise description of the audio content. |
| `tag_caption` | Comma-separated tags describing sounds and events. |

Each record is uniquely identified by `(dataset, id)`.

## Released annotations

| Dataset | Annotation records |
|---|---:|
| `audioset` | 683,316 |
| `audioset_sl` | 84,920 |
| `audiostock` | 9,001 |
| `bbc_sound_effects` | 45,577 |
| `epic_sound` | 60,055 |
| `free_to_use_sounds` | 5,733 |
| `freesound` | 256,432 |
| `sonniss_game_effects` | 5,005 |
| `soundbible` | 1,232 |
| `vggsound` | 173,913 |
| **Total** | **1,325,184** |

For AudioCaps and Clotho annotations, please refer to the original official documentation:

- [AudioCaps](https://audiocaps.github.io/)
- [Clotho](https://zenodo.org/records/3490684)

## Loading

Load the complete collection:

```python
from datasets import load_dataset

dataset = load_dataset("Jazzcharles/AudioVerse", "all", split="train")
```

Load an individual source using its name from the table above:

```python
dataset = load_dataset("Jazzcharles/AudioVerse", "sonniss_game_effects", split="train")
```