CD-ADD

CD-ADD数据集由华为翻译服务中心创建，包含超过300小时的语音数据，由五种先进的零样本TTS模型生成。数据集旨在模拟真实世界场景，采用多种攻击方法和来自不同数据集的音频提示。该数据集的创建过程涉及使用自动语音识别模型进行质量控制，确保生成的语音满足一定的字符错误率阈值。CD-ADD数据集主要用于跨域音频深度伪造检测，旨在解决随着零样本TTS技术发展而日益严重的个人权利和隐私侵犯问题。The CD-ADD dataset was created by the Huawei Translation Service Center. It comprises over 300 hours of speech data generated by five state-of-the-art zero-shot TTS models. The dataset is designed to simulate real-world scenarios, incorporating multiple attack methods and audio prompts derived from various datasets. During its development, quality control was carried out using automatic speech recognition (ASR) models to ensure the generated speech meets a predefined Character Error Rate (CER) threshold. The CD-ADD dataset is primarily used for cross-domain audio deepfake detection, aiming to address the increasingly severe personal rights and privacy infringement problems exacerbated by the advancement of zero-shot TTS technologies.

搜集汇总

背景与挑战

背景概述

CD-ADD数据集由华为翻译服务中心创建，包含超过300小时由五种零样本TTS模型生成的语音数据，旨在模拟真实场景以支持跨域音频深度伪造检测，通过ASR质量控制确保数据质量，应对零样本TTS技术发展带来的隐私和安全挑战。