Bilibili Video Analysis

Problem

Bilibili videos usually separate the visual track, audio track, and subtitles. Manually clipping clips, transcribing speech, and summarizing frames is fragmented. This skill targets short-video analysis and turns a link into a readable, reviewable summary.

How It Works

• Input: accepts short links, full Bilibili URLs, or raw BVID values.
• Download and preprocessing: resolves download URLs with bilibili-api-python, downloads m4s streams with aiohttp, converts audio to 16kHz mono WAV, and extracts keyframes with ffmpeg.
• Subtitles: runs local openai-whisper with the medium model to produce Chinese time-stamped subtitles.
• Frame analysis: selects 6-11 frames based on duration and analyzes them with the image tool.
• Summary: combines subtitles, frames, and visual notes into a structured summary, then sends it to Feishu via the message tool.

Boundaries and Notes

The skill depends on ffmpeg, a cached Whisper model, and local storage, and is aimed at short videos; a 60-second clip takes roughly 3-4 minutes. Downloads require a Referer, and file names use BVID to avoid special characters. Whisper may output Traditional Chinese; conversion is optional. It does not make copyright determinations and does not guarantee stable access to every Bilibili video.