DFADD
收藏资源简介:
DFADD数据集是由国立台湾大学创建的,专注于基于扩散和流匹配技术的音频深度伪造数据集。该数据集包含163,500条音频样本,涵盖了109个不同的说话者,每个说话者对应300个句子。数据集的创建过程包括文本选择和语音合成两个阶段,使用了五种不同的扩散和流匹配TTS模型。DFADD数据集主要用于开发和评估反欺骗模型,旨在解决当前反欺骗模型在面对高级扩散和流匹配TTS系统生成的音频时的不足。
The DFADD dataset, created by National Taiwan University, is a specialized audio deepfake dataset focused on diffusion and flow matching technologies. It contains 163,500 audio samples across 109 distinct speakers, with 300 sentences per speaker. The dataset construction consists of two stages: text selection and speech synthesis, and it leverages five different diffusion and flow matching-based TTS models. The DFADD dataset is primarily intended for developing and evaluating anti-spoofing models, with the goal of addressing the shortcomings of current anti-spoofing models when confronted with audio generated by state-of-the-art diffusion and flow matching TTS systems.
DFADD: The Diffusion and Flow-matching based Audio Deepfake Dataset
关键特性
- DFADD是首个包含使用扩散和Flow-matching技术生成的伪造语音的数据集。
- 与在ASVspoof数据集上训练的反伪造模型相比,在DFADD上训练的模型在面对相同方法生成的伪造语音时表现出更好的等误差率(EERs)。
数据集下载
-
HuggingFace数据集 python from datasets import load_dataset DFADD = load_dataset(isjwdu/DFADD)
致谢
DFADD的创建基于以下开源实现和数据集:




