QinboZhang/AudioSet

--- 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 configs: - config_name: balanced default: true data_files: - split: train path: data/bal_train/*.parquet - split: test path: data/eval/*.parquet - config_name: unbalanced data_files: - split: train path: data/unbal_train/*.parquet - split: test path: data/eval/*.parquet - config_name: full data_files: - split: bal_train path: data/bal_train/*.parquet - split: unbal_train path: data/unbal_train/*.parquet - split: eval path: data/eval/*.parquet --- # 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|18683|17141| #### `unbalanced` configuration | |train |test | |-----------|------:|----:| |# instances|1738657|17141| ## 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: 18683 audio clips out of 22160 originally. - Unbalanced train: 1738788 clips out of 2041789 originally. - Evaluation: 17141 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} } ```语言： - 英语 许可协议：知识共享署名4.0协议（CC BY 4.0） 规模类别： - 10K < 样本数量 < 100K - 1M < 样本数量 < 10M 源数据集： - 原创数据集 任务类别： - 音频分类 PapersWithCode编号：audioset 易读名称：AudioSet 配置名称： - 平衡版 - 非平衡版 标签： - 音频 配置项： - 配置名称：平衡版 默认启用：是 数据文件： - 拆分集：训练集 路径：data/bal_train/*.parquet - 拆分集：测试集 路径：data/eval/*.parquet - 配置名称：非平衡版 数据文件： - 拆分集：训练集 路径：data/unbal_train/*.parquet - 拆分集：测试集 路径：data/eval/*.parquet - 配置名称：完整版 数据文件： - 拆分集：平衡训练集 路径：data/bal_train/*.parquet - 拆分集：非平衡训练集 路径：data/unbal_train/*.parquet - 拆分集：评估集 路径：data/eval/*.parquet # AudioSet 数据集卡片 ## 数据集描述 - **主页**：https://research.google.com/audioset/index.html - **论文**：https://storage.googleapis.com/gweb-research2023-media/pubtools/pdf/45857.pdf - **排行榜**：https://paperswithcode.com/sota/audio-classification-on-audioset ### 数据集概述 [AudioSet](https://research.google.com/audioset/dataset/index.html) 是一个源自YouTube的10秒音频片段数据集，依据AudioSet本体论为每个片段标注一个或多个声音类别。 ### 支持任务与排行榜 - `audio-classification`：将音频片段分类至对应类别。排行榜详见[此处](https://paperswithcode.com/sota/audio-classification-on-audioset)。 ### 语言 数据集中的类别标签均为英语。 ## 数据集结构 ### 数据实例 数据集的示例实例如下： 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': ['语音', '喷涌声'] } ### 数据字段 数据实例包含以下字段： - `video_id`：字符串类型特征，存储原始YouTube视频的唯一标识符。 - `audio`：音频特征，包含音频数据与采样率信息。 - `labels`：字符串序列特征，存储音频片段对应的标签编码。 - `human_labels`：字符串序列特征，存储`labels`中标签对应的人类可读名称。 ### 数据拆分 音频片段的分布情况如下： #### 平衡版配置 | | 训练集 | 测试集 | |-----------|-------:|-------:| | 实例数量 | 18683 | 17141 | #### 非平衡版配置 | | 训练集 | 测试集 | |-----------|-----------:|-------:| | 实例数量 | 1738657 | 17141 | ## 数据集构建 ### 遴选依据 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ### 源数据 #### 初始数据收集与归一化 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) #### 源标签生产者来自何处？ 标签源自AudioSet本体论，音频片段均采集自YouTube平台。 ### 标注 #### 标注流程 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) #### 标注人员来自何处？ [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ### 个人与敏感信息 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ## 数据使用注意事项 ### 数据集的社会影响 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ### 偏差讨论 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ### 其他已知局限性 1. 本副本中的YouTube视频于2023年3月完成下载，因此并非所有原始音频片段均可获取。可下载的片段数量如下： - 平衡训练集：原始共22160条，现可获取18683条音频片段。 - 非平衡训练集：原始共2041789条，现可获取1738788条音频片段。 - 评估集：原始共20371条，现可获取17141条音频片段。 2. 多数音频采用48 kHz、24位采样格式，约10%的音频采用44.1 kHz、24位采样格式。音频文件以FLAC（Free Lossless Audio Codec，无损音频编码）格式存储。 ## 附加信息 ### 数据集遴选者 [需补充更多信息](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards) ### 许可信息 AudioSet数据集采用知识共享署名4.0协议（CC BY 4.0）进行授权。 ## 引用 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} }