Famous Figures Dataset
收藏资源简介:
Famous Figures数据集由美国密歇根大学电气与计算机工程系创建,旨在为政治人物提供高质量的语音合成数据,用于研究和开发音频反欺骗检测系统。数据集包含10位知名政治人物的26,500条真实语音样本和265,000条合成语音样本,平均时长约为8秒。数据集的创建过程涉及从YouTube视频收集高质量音频,使用Assembly AI进行说话人分割,OpenAI Whisper Large Turbo进行转录,并基于转录进行音频分割。合成语音生成采用了多种文本到语音(TTS)系统,包括单说话人模型、少量样本微调和零样本合成。该数据集可用于解决音频欺骗检测问题,保护公众人物免受语音伪造攻击。
The Famous Figures dataset was created by the Department of Electrical and Computer Engineering, University of Michigan, United States. It aims to provide high-quality speech synthesis data for political figures to support research and development of audio anti-spoofing detection systems. The dataset contains 26,500 real speech samples and 265,000 synthetic speech samples from 10 well-known political figures, with an average duration of approximately 8 seconds per sample. The dataset development process involves collecting high-quality audio from YouTube videos, performing speaker diarization via Assembly AI, conducting transcription using OpenAI Whisper Large Turbo, and executing audio segmentation based on the resulting transcriptions. For synthetic speech generation, multiple text-to-speech (TTS) systems are utilized, including single-speaker models, few-shot fine-tuning, and zero-shot synthesis. This dataset can be used to address audio spoofing detection challenges and protect public figures against voice forgery attacks.
数据集概述
ASVSpoof Laundered
- 数据集类型:未明确说明
- 应用领域:可能与音频欺骗检测相关(基于实验室研究方向推断)
- 实验室研究背景:信号处理、大数据分析、深度伪造检测
Famous Figures
- 数据集类型:未明确说明
- 应用领域:可能与人物识别或深度伪造检测相关(基于实验室研究方向推断)
- 实验室研究背景:多媒体取证、信息安全性
实验室资源支持
- 计算资源:高性能计算集群
- 设备支持:现代智能手机阵列、各类传感器
- 专项研究设备:福特福克斯车辆(用于汽车网络安全研究)

- 1Collecting, Curating, and Annotating Good Quality Speech deepfake dataset for Famous Figures: Process and Challenges美国密歇根大学电气与计算机工程系 · 2025年



