InTheWild
收藏资源简介:
In-the-Wild音频深度伪造数据集是一个专为语音反欺骗和合成语音检测任务设计的基准测试就绪数据集,由Müller等人在2022年的研究(arXiv:2203.16263)中提出,旨在评估音频深度伪造检测模型在真实世界条件下的跨域泛化能力。该数据集包含从公开来源收集的政治家和公众人物的真实语音(bonafide)与音频深度伪造(spoof)配对,用于二分类任务,总共有31,779个音频片段,其中19,963个为真实语音,11,816个为伪造语音。所有音频均为16 kHz单声道格式,以FLAC编码存储,并以规范的Parquet文件格式提供,包含四个字段:path(音频文件路径字符串)、audio(Audio(16000)对象)、label(分类标签,取值为bonafide或spoof)和notes(JSON字符串,包含唯一的utterance_id、speaker名称和源label字符串)。该数据集适用于音频分类任务,特别是作为跨域泛化基准,用于测试在实验室数据集(如ASVspoof)上训练的模型在真实场景中的性能。
The In-the-Wild Audio Deepfake Dataset is a benchmark-ready dataset designed for voice anti-spoofing and synthetic speech detection tasks. It was proposed by Müller et al. in 2022 (arXiv:2203.16263) to evaluate the cross-domain generalization capabilities of audio deepfake detection models under real-world conditions. The dataset contains bonafide (real) and spoof (audio deepfake) pairs collected from public sources of politicians and public figures for binary classification tasks, with a total of 31,779 audio clips, including 19,963 bonafide and 11,816 spoof clips. All audio is in 16 kHz mono format, encoded in FLAC, and provided in a standardized Parquet file format with four fields: path (audio file path string), audio (Audio(16000) object), label (classification label, with values bonafide or spoof), and notes (JSON string containing unique utterance_id, speaker name, and source label string). The dataset is suitable for audio classification tasks, particularly as a cross-domain generalization benchmark to test the performance of models trained on laboratory datasets (e.g., ASVspoof) in real-world scenarios.
数据集概述
- 数据集名称:In-the-Wild Audio Deepfake Dataset
- 许可证:Apache License 2.0
- 语言:英语
- 任务类别:音频分类
- 数据集规模:10K < n < 100K(共31,779个音频片段)
- 相关论文:arXiv 2203.16263
数据集描述
In-the-Wild是一个用于语音反欺骗/合成语音检测的音频深度伪造数据集,旨在评估模型在跨域泛化场景下的性能。数据集包含政治家和公众人物的真实语音与音频深度伪造样本,所有样本均来自公开可用的来源。
任务与标签
- 任务类型:二分类
- bonafide(真实人类语音):19,963个片段
- spoof(深度伪造语音):11,816个片段
数据格式
- 音频格式:16 kHz单声道FLAC编码(16位PCM)
- 数据文件:Parquet格式,包含4列:
path:字符串,音频文件路径audio:Audio(16000),音频数据label:ClassLabel[bonafide, spoof],类别标签notes:JSON字符串,包含utterance_id(唯一标识)、speaker(说话人名称)、label(源标签字符串)
数据划分
数据集仅提供测试集(test split),数据文件路径为 data/test-*.parquet。
引用
bibtex @inproceedings{muller2022does, title={Does Audio Deepfake Detection Generalize?}, author={M{"u}ller, Nicolas M and Czempin, Pavel and Dieckmann, Franziska and Froghyar, Adam and B{"o}ttinger, Konstantin}, booktitle={Interspeech}, year={2022} }




