CodecDeepfakeDetection (CDD)
收藏资源简介:
CodecDeepfakeDetection是由斯图加特大学和AppTek联合构建的高挑战性音频深度伪造检测数据集,作为ASVspoof 5的扩展版本。该数据集包含36,000条样本,覆盖训练集、开发集和测试集,数据来源包括真实语音、基于神经音频编解码器(如EnCodec、Mimi等)重合成的音频,以及6种先进文本转语音系统生成的伪造音频。数据集通过严格划分说话人重叠比例和平衡编解码器类型,模拟了真实场景中编解码器在传输压缩与语音合成中的双重应用。其核心价值在于为解决编解码重合成音频的标签歧义问题提供基准,推动音频伪造检测模型在复杂场景下的鲁棒性研究。
CodecDeepfakeDetection is a highly challenging audio deepfake detection dataset jointly developed by the University of Stuttgart and AppTek, serving as an extended version of ASVspoof 5. This dataset consists of 36,000 samples across training, development, and test splits, with data sources encompassing real speech, resynthesized audio generated via neural audio codecs (e.g., EnCodec, Mimi, etc.), and forged audio produced by six state-of-the-art text-to-speech systems. By strictly regulating the speaker overlap ratio and balancing codec types, the dataset simulates the dual application of audio codecs in transmission compression and speech synthesis in real-world scenarios. Its core value lies in providing a benchmark for addressing the label ambiguity problem of codec-resynthesized audio, and facilitating research on the robustness of audio deepfake detection models in complex scenarios.
- 1How to Label Resynthesized Audio: The Dual Role of Neural Audio Codecs in Audio Deepfake Detection斯图加特大学·自然语言处理研究所; AppTek有限公司 · 2026年



