DeepFake-Audio-Rangers/Arabic_Audio_Deepfake
收藏资源简介:
--- dataset_info: features: - name: audio dtype: audio - name: label dtype: class_label: names: '0': fake '1': real splits: - name: train num_bytes: 1322891516.5301797 num_examples: 15648 - name: test num_bytes: 331978521.95082015 num_examples: 3913 download_size: 1567148799 dataset_size: 1654870038.481 configs: - config_name: default data_files: - split: train path: data/train-* - split: test path: data/test-* task_categories: - audio-to-audio - audio-classification - automatic-speech-recognition language: - ar tags: - synthetic size_categories: - 10K<n<100K pretty_name: ArAD license: odc-by --- ### ArAD Dataset (Arabic Audio DeepFake Dataset) **Dataset Summary** This dataset contains Arabic deepfake audio samples, focusing mainly on Levantine dialect with some examples in Standard Arabic. It was created using the RVC v2 framework, fine-tuned on a custom dataset of multi-dialect Arabic speech. The goal is to simulate real-world deepfake audio attacks by generating synthetic speech from actual recordings and voice messages. One of the first datasets to include real-world deepfake audio arabic speech. **Dataset Creation** The process involved collecting speech audio for each speaker (with a minimum of 5 minutes per speaker) and training a unique model for each speaker. These models were then used to generate fake audio, ensuring no speaker's model was used to generate their own voice. Speech clips were segmented based on silence and limited to a maximum of 3 seconds in length. **Audio Specifications** - Format: WAV - Sample Rate: 16KHz (resampled from original recordings at various sample rates) - Audio cleaning: Major disturbances were cut, audio was cleaned using Ultimate Vocal Remover or the Adobe Podcast platform. - Resampling: Audio was resampled to 40KHz for model generation, then downsampled to 16KHz using the Librosa Python library. **Intended Use** This dataset is valuable for research in deepfake detection, voice cloning, and related AI tasks. It is a resource for exploring vulnerabilities in voice systems. **Licensing and Citation** This data is made available under the Open Data Commons Attribution License: http://opendatacommons.org/licenses/by/1.0/
数据集信息: 特征: - 名称:音频(audio) 数据类型:音频(audio) - 名称:标签(label) 数据类型: 类别标签(class_label): 类别名称: '0': 伪造(fake) '1': 真实(real) 划分: - 名称:训练集(train) 字节数:1322891516.5301797 样本数:15648 - 名称:测试集(test) 字节数:331978521.95082015 样本数:3913 下载大小:1567148799 数据集总大小:1654870038.481 配置: - 配置名称:默认(default) 数据文件: - 划分:训练集(train) 路径:data/train-* - 划分:测试集(test) 路径:data/test-* 任务类别: - 音频到音频(audio-to-audio) - 音频分类(audio-classification) - 自动语音识别(automatic-speech-recognition) 语言: - 阿拉伯语(ar) 标签: - 合成(synthetic) 样本量类别: - 10K<n<100K 美观名称:ArAD 许可:odc-by(Open Data Commons Attribution License,开放数据共享署名许可) --- ### ArAD数据集(阿拉伯语音频深度伪造数据集) **数据集摘要** 本数据集包含阿拉伯语深度伪造音频样本,主要聚焦黎凡特方言,同时包含部分现代标准阿拉伯语示例。本数据集基于RVC v2框架构建,在多方言阿拉伯语语音自定义数据集上完成微调。其构建目标为通过真实录音与语音消息生成合成语音,模拟现实世界中的深度伪造音频攻击场景。本数据集是首批收录真实世界阿拉伯语深度伪造语音的数据集之一。 **数据集构建流程** 该数据集的构建过程包括:为每位说话人采集语音音频(每位说话人至少5分钟音频素材),并为每位说话人训练专属模型;随后使用这些模型生成伪造音频,且严格避免使用某一说话人的模型生成其自身的语音。语音片段按静音阈值进行分段,且单段最长时长限制为3秒。 **音频规格说明** - 格式:WAV - 采样率:16KHz(从多种原始采样率的录音文件重采样得到) - 音频清理:移除主要干扰信号,通过终极人声移除工具(Ultimate Vocal Remover)或Adobe Podcast平台完成音频降噪与清理。 - 重采样流程:生成模型所用音频先被重采样至40KHz,随后通过Librosa Python库下采样至16KHz。 **预期用途** 本数据集可用于深度伪造检测、语音克隆及相关人工智能任务的研究,是探索语音系统安全脆弱性的宝贵研究资源。 **许可与引用说明** 本数据集基于开放数据共享署名许可(Open Data Commons Attribution License)发布,许可详情链接:http://opendatacommons.org/licenses/by/1.0/




