FaizanMirza123/talkingcelebs

TalkingCelebs是一个手动策划的视频数据集，专门用于说话动作签名（Talking Motion Signatures, TMS）分析和深度伪造检测研究。该数据集包含500个高质量视频片段，来自五位知名公众人物（巴拉克·奥巴马、安格拉·默克尔、弗拉基米尔·泽连斯基、埃隆·马斯克、艾玛·沃特森），每个身份100个片段，每个片段时长为30秒，包含连续语音。数据集旨在解决现有数据集的局限性，提供更多每个身份的片段和更长的持续时间片段。视频来源为YouTube上的公开演讲、采访和演示，每个身份有13-18个独特视频。分辨率统一为720p，帧率为25 FPS。数据集按80%训练集（400个片段）和20%测试集（100个片段）划分，确保训练和测试视频来源无重叠。数据经过预处理，包括帧重缩放至720p和标准化至25 FPS，并经过手动质量验证。数据集适用于深度伪造检测、生物特征认证和视频分析研究，但需负责任使用，避免用于创建深度伪造或误导性内容。数据集在人口统计学和语境上存在局限性，仅包含五位个体，且主要为正式演讲语境下的英语内容。TalkingCelebs is a manually curated video dataset designed specifically for Talking Motion Signatures (TMS) analysis and deepfake detection research. The dataset contains 500 high-quality video clips from five prominent public figures (Barack Obama, Angela Merkel, Volodymyr Zelenskyy, Elon Musk, Emma Watson), with 100 clips per identity, each clip lasting 30 seconds and containing continuous speech. It addresses limitations of existing datasets by providing more clips per identity and longer duration clips. Video sources are from publicly available YouTube videos of speeches, interviews, and presentations, with 13-18 unique videos per person. Resolution is standardized to 720p, frame rate to 25 FPS. The dataset is split into 80% training (400 clips) and 20% test (100 clips), with no overlap in source videos between splits. Data preprocessing includes frame rescaling to 720p and standardization to 25 FPS, with manual quality control. Intended uses include research on deepfake detection, biometric authentication, and video analysis, but it should be used responsibly, not for creating deepfakes or misleading content. The dataset has biases in representation (limited to 5 individuals) and context (primarily formal English-language speaking contexts).