VocalNet/VocalBench

VocalBench是一个用于评估语音交互模型在语音对话能力方面的基准测试数据集。它包含多种任务类别，如问答、音频转换、音频理解等，并支持多种语言。数据集包含多个配置，每个配置都有不同的特征和数据分割方式。例如，emotion配置包含问题ID、音频、问题文本、情感标签和评分等特征，并且只包含测试分割。single_round配置包含问题ID、音频、问题文本、类别和来源等特征，同样只包含测试分割。其他配置还包括creativity、knowledge、instruction、reasoning、safety、multi_round和robust等。

VocalBench is a benchmark dataset for evaluating the vocal conversational abilities of speech interaction models. It includes various task categories such as question-answering, audio-to-audio, and audio understanding, and supports multiple languages. The dataset contains multiple configurations, each with different features and data splitting methods. For example, the emotion configuration includes features such as question ID, audio, question text, emotion label, and score, and only contains the test split. The single_round configuration includes features such as question ID, audio, question text, category, and source, and also only contains the test split. Other configurations include creativity, knowledge, instruction, reasoning, safety, multi_round, and robust.