ADU-Bench
收藏资源简介:
ADU-Bench是一个用于评估大型音频语言模型(LALMs)在开放式音频对话理解能力的综合基准数据集。该数据集由清华大学、牛津大学和鹏城实验室联合创建,包含20,715条开放式音频对话,涵盖了多种语言和技能领域。数据集的创建过程结合了真实世界录音和合成音频样本,旨在测试LALMs在处理数学符号、理解人类行为、多语言对话以及处理音频对话中的歧义等方面的能力。ADU-Bench的应用领域广泛,旨在解决LALMs在实际应用中遇到的音频对话理解问题,特别是在多语言和多场景下的对话处理。
ADU-Bench is a comprehensive benchmark dataset for evaluating Large Audio-Language Models (LALMs) on open-ended audio dialogue understanding. This dataset was jointly developed by Tsinghua University, the University of Oxford, and the Peng Cheng Laboratory, and comprises 20,715 open-ended audio dialogue samples covering diverse languages and skill domains. The dataset’s creation process integrates real-world recordings and synthetic audio samples, aiming to test LALMs’ capabilities in handling mathematical notations, comprehending human behaviors, engaging in multilingual dialogues, and resolving ambiguities within audio conversations. ADU-Bench has a wide range of application scenarios, aiming to address audio dialogue understanding challenges encountered by LALMs in real-world applications, particularly dialogue processing across multilingual and multi-scenario settings.




