AudioSet / README.md
agkphysics's picture
Add unbalanced config. Fix memory leak.
5a2fa42 unverified
---
language:
- en
license: cc-by-4.0
size_categories:
- 10K<n<100K
- 1M<n<10M
source_datasets:
- original
task_categories:
- audio-classification
paperswithcode_id: audioset
pretty_name: AudioSet
config_names:
- balanced
- unbalanced
tags:
- audio
dataset_info:
- config_name: balanced
features:
- name: video_id
dtype: string
- name: audio
dtype: audio
- name: labels
sequence: string
- name: human_labels
sequence: string
splits:
- name: train
num_bytes: 26016210987
num_examples: 18685
- name: test
num_bytes: 23763682278
num_examples: 17142
download_size: 49805654900
dataset_size: 49779893265
- config_name: unbalanced
features:
- name: video_id
dtype: string
- name: audio
dtype: audio
- name: labels
sequence: string
- name: human_labels
sequence: string
splits:
- name: train
num_bytes: 2408656417541
num_examples: 1738788
- name: test
num_bytes: 23763682278
num_examples: 17142
download_size: 2433673104977
dataset_size: 2432420099819
---
# Dataset Card for AudioSet
## Dataset Description
- **Homepage**: https://research.google.com/audioset/index.html
- **Paper**: https://storage.googleapis.com/gweb-research2023-media/pubtools/pdf/45857.pdf
- **Leaderboard**: https://paperswithcode.com/sota/audio-classification-on-audioset
### Dataset Summary
[AudioSet](https://research.google.com/audioset/dataset/index.html) is a
dataset of 10-second clips from YouTube, annotated into one or more
sound categories, following the AudioSet ontology.
### Supported Tasks and Leaderboards
- `audio-classification`: Classify audio clips into categories. The
leaderboard is available
[here](https://paperswithcode.com/sota/audio-classification-on-audioset)
### Languages
The class labels in the dataset are in English.
## Dataset Structure
### Data Instances
Example instance from the dataset:
```python
{
'video_id': '--PJHxphWEs',
'audio': {
'path': 'audio/bal_train/--PJHxphWEs.flac',
'array': array([-0.04364824, -0.05268681, -0.0568949 , ..., 0.11446512,
0.14912748, 0.13409865]),
'sampling_rate': 48000
},
'labels': ['/m/09x0r', '/t/dd00088'],
'human_labels': ['Speech', 'Gush']
}
```
### Data Fields
Instances have the following fields:
- `video_id`: a `string` feature containing the original YouTube ID.
- `audio`: an `Audio` feature containing the audio data and sample rate.
- `labels`: a sequence of `string` features containing the labels
associated with the audio clip.
- `human_labels`: a sequence of `string` features containing the
human-readable forms of the same labels as in `labels`.
### Data Splits
The distribuion of audio clips is as follows:
#### `balanced` configuration
| |train|test |
|-----------|----:|----:|
|# instances|18685|17142|
#### `unbalanced` configuration
| |train |test |
|-----------|------:|----:|
|# instances|1738788|17142|
## Dataset Creation
### Curation Rationale
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
### Source Data
#### Initial Data Collection and Normalization
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
#### Who are the source language producers?
The labels are from the AudioSet ontology. Audio clips are from YouTube.
### Annotations
#### Annotation process
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
#### Who are the annotators?
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
### Personal and Sensitive Information
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
## Considerations for Using the Data
### Social Impact of Dataset
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
### Discussion of Biases
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
### Other Known Limitations
1. The YouTube videos in this copy of AudioSet were downloaded in March
2023, so not all of the original audios are available. The number of
clips able to be downloaded is as follows:
- Balanced train: 18685 audio clips out of 22160 originally.
- Unbalanced train: 1738788 clips out of 2041789 originally.
- Evaluation: 17142 audio clips out of 20371 originally.
2. Most audio is sampled at 48 kHz 24 bit, but about 10% is sampled at
44.1 kHz 24 bit. Audio files are stored in the FLAC format.
## Additional Information
### Dataset Curators
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
### Licensing Information
The AudioSet data is licensed under CC-BY-4.0
## Citation
```bibtex
@inproceedings{jort_audioset_2017,
title = {Audio Set: An ontology and human-labeled dataset for audio events},
author = {Jort F. Gemmeke and Daniel P. W. Ellis and Dylan Freedman and Aren Jansen and Wade Lawrence and R. Channing Moore and Manoj Plakal and Marvin Ritter},
year = {2017},
booktitle = {Proc. IEEE ICASSP 2017},
address = {New Orleans, LA}
}
```