Audio Dialogues

Audio Dialogues是由英伟达创建的一个多轮对话数据集，专注于音频和音乐理解，包含163.8k样本。该数据集利用大型语言模型（LLM）通过提示方法生成多轮对话，并从现有的音频数据集中提取描述性注释。数据集内容涵盖一般音频声音和音乐，支持复杂交互，如使用代词和基于先前回答的后续问题。创建过程中，采用了数据过滤策略以确保对话质量。该数据集适用于训练和评估音频增强的大型语言模型，旨在提高模型在音频相关领域的交互能力和理解深度。

Audio Dialogues is a multi-turn dialogue dataset developed by NVIDIA, focused on audio and music understanding, with a total of 163.8k samples. This dataset generates multi-turn dialogues via prompting-based methods using Large Language Models (LLMs), while extracting descriptive annotations from existing audio datasets. The dataset covers general audio sounds and music, and supports complex interactive scenarios such as employing pronouns and follow-up questions grounded in prior conversational responses. Data filtering strategies were adopted throughout the construction process to ensure the quality of the dialogues. This dataset is suitable for training and evaluating audio-augmented large language models, aiming to enhance the interactive capabilities and in-depth comprehension performance of models across audio-related fields.