Edited Audio Datasets (EADs)
收藏资源简介:
Edited Audio Datasets (EADs) 是由香港科技大学(广州)等机构创建的一个音频数据集,旨在评估音频编辑对大型音频语言模型(LALMs)的影响。该数据集包含520条音频样本,这些样本是通过将AdvBench中的有害文本问题转换为音频生成的。数据集的内容涵盖了多种音频编辑方法,如音调调整、词语强调、噪声注入等。数据集的创建过程包括使用gTTS将文本转换为音频,并应用多种音频编辑技术生成多样化的音频样本。该数据集的应用领域主要集中在音频语言模型的安全性研究,旨在解决音频编辑对模型推理输出的影响问题,特别是模型在面对音频编辑时的鲁棒性和安全性。
Edited Audio Datasets (EADs) is an audio dataset developed by institutions including The Hong Kong University of Science and Technology (Guangzhou), which aims to evaluate the impact of audio editing on Large Audio Language Models (LALMs). The dataset consists of 520 audio samples generated by converting harmful textual questions from the AdvBench benchmark into audio. Its content covers a wide range of audio editing techniques, such as pitch adjustment, word emphasis, noise injection, and more. The dataset creation workflow includes converting text to audio via gTTS, followed by applying various audio editing technologies to produce diverse audio samples. The primary application field of this dataset lies in the safety research of audio language models, targeting the investigation of the impact of audio editing on model inference outputs, particularly the robustness and safety of such models when confronted with edited audio.

- 1Tune In, Act Up: Exploring the Impact of Audio Modality-Specific Edits on Large Audio Language Models in Jailbreak香港科技大学(广州)、牛津大学、东北大学、西安交通大学、德雷克塞尔大学 · 2025年



