slprl/common-sense-facts-audio
收藏资源简介:
该数据集包含三种配对版本的语音常识事实提示:1. prompt:不完整的事实提示,例如“法国的首都是”;2. fact:正确的完整句子,例如“法国的首都是巴黎”;3. counterfactual:错误的完整句子,例如“法国的首都是罗马”。数据集未定义训练/验证/测试分割,所有示例都包含在名为“data”的单一中性分割中。每个示例对应一个基本事实,包含三个音频文件(prompt_audio、fact_audio、counterfactual_audio)以及提示音频的单词级时间戳(prompt_words、prompt_word_starts、prompt_word_ends)。数据集共有281个示例,涵盖多个类别,如颜色、星期、月份、对象功能、常识事实、国家语言、家庭关系、数值事实、反义词、婴儿动物与职业、首都城市、简单算术和数字序列。音频文件为WAV格式,采样率可能不同,时间戳通过强制对齐获得。数据集旨在用于语音-语言模型对齐、口语语言模型中的事实回忆和探测研究。
This dataset contains spoken common-sense factual prompts in three paired versions: 1. prompt: incomplete factual prompt, e.g., the capital of france is; 2. fact: correct full sentence, e.g., the capital of france is paris; 3. counterfactual: incorrect full sentence, e.g., the capital of france is rome. The dataset does not define train/validation/test splits; all examples are provided in a single neutral split named data. Each row corresponds to one base fact and contains three audio files (prompt_audio, fact_audio, counterfactual_audio) and word-level timestamps for the prompt audio (prompt_words, prompt_word_starts, prompt_word_ends). There are 281 examples covering categories such as Colors, Days of the week, Months of the year, Object functions, Common-sense facts, Country languages, Family relations, Numerical facts, Opposites, Baby animals & professions, Capital cities, Simple arithmetic, and Number sequences. Audio files are in WAV format with varying sampling rates, and timestamps are obtained via forced alignment. The dataset is intended for research on speech-language model alignment, factual recall in spoken language models, and probing studies.




