AI Subtitle Extractor

August 22, 2026 · View on GitHub

Author website · X · Bilibili · YouTube · Sponsor

Turn any online video link into a clean transcript — read the subtitles the platform already has. Human tracks first, auto tracks as fallback. No video download, no ASR.

The core of this repo is the recipe in SKILL.md: one shared pipeline, with YouTube and Bilibili as verified adapters. Any other site falls back to generic discovery — no support claimed.

video link
  → confirm lawful use + exact video + operation surface
  → identify site (or generic discovery)
  → ask the target language first — translate into the language the user wants
  → find captions (API / timedtext / VTT·SRT / transcript panel / <track>)
  → human track first → capture and verify all timed cues
  → remove speech noise while preserving every substantive claim and detail
  → restructure as a natural article and fully translate when needed
  → deliver the complete article before summaries, Q&A, or other follow-up work
  → site unreachable: fall back to the user's local browser
  → truly no captions: say so (only then consider ASR)

Verification status

Only what actually ran is claimed.

Component	Status
YouTube adapter	✅ Verified on a real watch page (via WebBridge)
Bilibili adapter	✅ Verified on real videos (full SRT/text export)
Generic sniffing (fetch/XHR hook, `textTracks`, `<track>`)	✅ Verified in test suite + real YouTube page
Injection via WebBridge `evaluate` (main path)	✅ Verified
Injection via Playwright `add_init_script`	✅ Verified (whole test suite)
MV3 extension form	✅ Verified on Bilibili
Tampermonkey install form	⚠️ Not independently tested
agent-browser backend in `scripts/`	⚠️ Not yet tested
Acknowledged download + complete source/cue return	✅ Verified in Chromium mock
Source-cue capture integrity	✅ Verified in pure Python + Chromium mock
Target-scoped YouTube “Skip ad” click	✅ Verified in Chromium mock; real ads pending
Any other site	Generic discovery only

Quick start (verified main path)

Drive the user's real browser (their login session) through a bridge like Kimi WebBridge — nothing to install:

1. Confirm lawful use, the exact video URL, and the operation surface with the user
2. node universal/build.js   → dist/universal-subtitle-extractor.user.js
3. navigate → the confirmed video page
4. evaluate → inject the .user.js file (re-injection guarded)
5. evaluate → bind player control to that exact target, then export:

(async () => {
  const videoUrl = location.href; // must equal the URL the user confirmed
  const auth = window.__USE__.authorizeTarget(videoUrl, {
    acknowledgeLawfulUse: true
  });
  if (!auth.ok) return auth;
  await window.__USE__.waitFor(1, 20000);
  return window.__USE__.download('txt', null, null, {
    targetUrl: videoUrl,
    acknowledgeLawfulUse: true
  });
})()

The page returns the complete source transcript, cues, and capture report with requiresEditorialPass; this is evidence, not the final article. The Agent removes filler, false starts, and meaningless repetition while retaining substantive claims, examples, numbers, conditions, and conclusions, then restructures and translates the complete article before follow-up work. Ad skipping is authorized only for that video id and stops after a YouTube SPA navigation to another video.

Repository layout

Path	What it is
`SKILL.md`	The recipe (core asset): pipeline → adapters → fallback → output contract
`universal/`	Runnable universal extractor (`window.__USE__` API), tests included
`scripts/`	Reference CLI + page-inject core (`__ovsExportSubtitle`)
`examples/`	Adapter walkthroughs: YouTube + browser fallback, Bilibili curl

Contributing

Site adapters welcome: follow the shared pipeline and the Cue model, use public examples, label "verified" or "experimental" honestly. See CONTRIBUTING.md.

Website and other links

No separate product site is required for this repository. The public face of the work is the author website, this GitHub repo, and the projects below.


Project page	https://guojiz.github.io/ai-subtitle-extractor/
Author website	https://guojiz.github.io/
X	https://x.com/guojizh
Bilibili	https://space.bilibili.com/3493114115263006
YouTube	https://youtube.com/@guojizh
Sponsor	https://github.com/Guojiz/Sponsors

Other open-source projects

GitLearnOS — learner-owned Git memory
Word Snap — bilingual vocabulary matching
AI Subtitle Extractor
Design Master
AI Video Studio
llm-provider-compat
Claude Desktop Tweak Models
All projects: github.com/Guojiz

License

MIT — see LICENSE.