Audio Dialogues
收藏资源简介:
Audio Dialogues是英伟达构建的一个多轮对话数据集,旨在提升音频和音乐理解能力。数据集包含163.8K个样本,涵盖通用声音和音乐,除了对话外,该数据集还包括问题-答案对,以理解和比较多个输入音频。该数据集通过基于提示的方法和大型语言模型生成,研究人员首先从AudioSet和MusicCaps这两个数据集中获取音频的注释,然后使用GPT-4模型根据这些注释生成多轮对话。Audio Dialogues的推出,为音频理解领域的研究提供了宝贵的资源,可用于训练和评估音频增强的大型语言模型。
Audio Dialogues is a multi-turn dialogue dataset constructed by NVIDIA to enhance audio and music understanding capabilities. It comprises 163.8K samples covering general sound and music domains. Beyond dialogues, the dataset also provides question-answer pairs to facilitate the understanding and comparison of multiple input audios. Generated via prompt-based approaches and large language models, this dataset is developed as follows: researchers first retrieve audio annotations from two existing datasets, AudioSet and MusicCaps, then leverage the GPT-4 model to generate multi-turn dialogues based on these annotations. The release of Audio Dialogues offers a valuable resource for audio understanding research, enabling the training and evaluation of audio-augmented large language models.




