Codecfake
收藏资源简介:
Codecfake数据集是由中国科学院自动化研究所创建的大型开放源代码数据集,专注于基于音频语言模型(ALM)的音频检测。该数据集包含超过100万音频样本,涵盖两种语言,并设置了多种测试条件,以评估模型对ALM生成音频的检测能力。数据集的构建涉及多种主流的神经音频编解码模型,旨在通过这些模型生成的伪造音频来训练和验证音频深度伪造检测(ADD)模型。Codecfake数据集的应用领域主要集中在提高对ALM生成音频的检测准确性,解决现有模型在检测新型伪造音频方面的局限性。
The Codecfake dataset is a large open-source dataset developed by the Institute of Automation, Chinese Academy of Sciences, focusing on audio detection based on Audio Language Models (ALM). This dataset contains over 1 million audio samples covering two languages, and is configured with multiple test conditions to evaluate models' performance in detecting audios generated by ALMs. The construction of the Codecfake dataset involves a variety of mainstream neural audio codec models, aiming to train and validate Audio Deepfake Detection (ADD) models using the forged audios generated by these models. The primary application of the Codecfake dataset is to improve the detection accuracy of ALM-generated audios and address the limitations of existing models in detecting novel forged audios.

- 1Codecfake: An Initial Dataset for Detecting LLM-based Deepfake Audio中国科学院自动化研究所 · 2024年



