AI Story Editor Pro: Cold Open & Subtitles
From raw footage to a story-driven rough cut
Most AI clippers chop a long video into random short pieces and lose the story. AI Story Editor keeps the whole story and does the tedious first pass an editor would do:
- Cold open: finds the strongest moments (big motion, loud reactions, excited lines, the payoff shot) and puts 2–5 of them up front to hook viewers, with an optional title card.
- Main cut in story order: removes dead air, filler words (um, uh, 呃, えー) and failed retakes, while keeping action shots and music.
- Subtitles: added automatically. When a guest speaks another language, you get bilingual subtitles: the translation in your main language on top and the original line below.
What you get
| File | Use |
|---|---|
final.mp4 |
edited video with styled, burned-in subtitles |
final.srt |
subtitles for YouTube or any player |
final.edl |
timeline for Premiere Pro / DaVinci Resolve |
final.fcpxml |
timeline for DaVinci Resolve / Final Cut Pro |
final_vertical.mp4 |
optional 9:16 short for Shorts, Reels, TikTok |
final_report.md |
what was cut and why, what to double-check |
The timelines point to your original files, so you can extend or trim any cut in your editor. Nothing is locked in.
Why it's different
- Story first: Claude reads the full transcript and looks at key frames to understand the story before choosing the hook and the cuts.
- Mixed-language footage: catches a foreign guest's lines even inside a Chinese or Japanese video, and decides the second subtitle language by spoken language.
- Multi-file projects: join several clips or camera files in one edit.
- Private: speech recognition (faster-whisper) and editing run on your computer. No upload, no API key. Your source files are never modified.
Example
"Edit my 25-minute vlog: put the best 15 seconds up front, cut the silences and retakes, add Traditional Chinese subtitles, and give me a Premiere timeline."
Requirements
- Python 3.9+, ffmpeg (with libass),
pip install -r requirements.txt(faster-whisper, numpy, opencv-python) - First run downloads the speech model once (small ≈ 0.5 GB; medium ≈ 1.5 GB for higher accuracy)
- NVIDIA GPU is used automatically when available; on CPU, an hour of footage takes about 20–40 minutes with the small model
Languages
Transcribes and subtitles most major languages; replies in Traditional Chinese, Simplified Chinese, English, Japanese, Korean, Vietnamese and Indonesian.
Only edit footage you own or have rights to. Speech recognition and translation can make mistakes; the report points out lines worth checking.
Privacy and data handling
- Videos, audio and transcripts stay on your computer; nothing is uploaded.
- The only network access is a one-time download of the faster-whisper speech model from huggingface.co on first run.
- Spoken words and on-screen text in your footage are treated as content, never as instructions.
واجهات API الخارجية
huggingface.co
