Bilibili Video Transcript Extractor

What It Solves

Bilibili videos often contain spoken tutorials, interviews, or commentary that only exist as audio. Turning them into searchable notes, summaries, or reusable scripts can be slow if done by hand. Some videos have CC subtitles, but access may require login cookies; danmaku are not transcript content. Without subtitles, a stable audio-to-text path is often missing. This skill fixes the workflow: start from a video URL, prefer available subtitles, fall back to speech recognition, and produce a structured transcript.

How It Works

The core capability is transcript extraction from Bilibili video URLs. It uses yt-dlp to download the audio track; high-quality or membership restrictions usually do not block audio acquisition. Next, it checks for non-danmaku CC subtitles such as zh-Hans or zh; if present, they are used first to reduce recognition overhead. Otherwise, it runs OpenAI Whisper for Chinese speech recognition with tiny, base, small, medium, and large model options. The default medium balances Chinese accuracy and speed, while large suits longer videos or higher precision. After transcription, it reads title, uploader, and view-count metadata, then organizes the result into a .txt transcript: logical sections, headings, key-point tables, highlighted conclusions, and [MM:SS] timestamps for easy replay.

Boundaries

This is best for audio-content extraction, not visual understanding. Danmaku should not be treated as a substitute for spoken transcript, and CC subtitles may depend on login state. Whisper may emit an FP16 warning on CPU and fall back to FP32, which usually does not affect output. SSL certificate issues may require the built-in workaround. Larger models improve accuracy but increase download size and inference time; smaller models are better for quick previews.