ronadin/ishowspeed-streams

该数据集包含来自IShowSpeed IRL直播流的229,959个场景级视觉描述和语音转录文本，覆盖656小时的直播内容。数据涵盖IShowSpeed的两个旗舰IRL巡演：Speed Does America（35天不间断直播，跨越美国25个州）和Speed Does Africa（30天，跨越非洲20个国家）。每个视频被分割为10秒窗口；对于每个窗口，数据集提供：由Gemini 2.5 Flash根据实际视频帧生成的1-3句视觉场景描述（描述相机运动、屏幕上的人物、位置、屏幕叠加内容和动作），以及从YouTube滚动自动字幕中提取并去重后的该窗口语音转录。数据集还包括视频ID、标题、巡演信息、YouTube链接、持续时间、场景索引、时间戳等元数据。The dataset contains 229,959 scene-level visual descriptions and spoken transcripts from 656 hours of IShowSpeeds IRL streams. It covers two of IShowSpeeds flagship IRL tours: Speed Does America (a 35-day non-stop livestream tour across 25 US states) and Speed Does Africa (a 30-day, 20-country tour across the African continent). Each video is split into 10-second windows; for every window, the dataset provides: a 1–3 sentence visual scene description generated by Gemini 2.5 Flash from the actual video frames (describing camera movement, whos on screen, location, on-screen overlays, and action), and the spoken transcript of that window, deduplicated from YouTubes rolling auto-captions. The dataset also includes metadata such as video ID, title, tour information, YouTube URL, duration, scene index, timestamps, etc.