WaveFake

该数据集名为Wavefake，它通过六种不同的声码器架构生成深度伪造音频，这些音频文件总共约196小时，源自LJSPEECH数据集。该数据集被划分为训练集、验证集和测试集，比例分别为70%、10%和20%。此外，该数据集能够对检测模型在干净和受损音频数据上进行评估。规模大约为196小时的音频，任务是对音频深度伪造进行检测。

This dataset, named Wavefake, generates deepfake audio using six distinct vocoder architectures. Derived from the LJSPEECH dataset, the total duration of its audio files is approximately 196 hours. It is split into training, validation, and test sets with a ratio of 70%, 10%, and 20% respectively. Additionally, this dataset can be used to evaluate detection models on both clean and corrupted audio data, with its core task being deepfake audio detection.

搜集汇总

背景与挑战

背景概述

WaveFake是一个专注于音频深度伪造检测的数据集，包含六种不同网络架构生成的样本，支持两种语言。该数据集提供了预训练模型和基线分类器，用于分析和检测音频深度伪造，旨在推动该领域的研究发展。