license: cc-by-4.0
task_categories:
- audio-classification
tags:
- audio
size_categories:
- 10K<n<100K
paperswithcode_id: audioset
dataset_info:
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
Dataset Card for AudioSet
Dataset Details
This repository contains the balanced training set and evaluation set of the AudioSet data. The YouTube videos were downloaded in March 2023, so not all of the original audios are available.
The distribuion of audio clips is as follows. In parentheses is the dict
key used for HugginFace datasets:
bal_train(train): 18685 audio clips out of 22160 originally.eval(test): 17142 audio clips out of 20371 originally.
You can use the datasets library to load this dataset, in which case
the raw audio will be returned along with a sequence of one or more
labels. Note that the raw audio is returned without further processing,
so you will need to decode and possibly downsample the audio for model
training.
Example instance from the train subset:
{
'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']
}
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.
Citation
@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}
}