An AI pipeline that extracts subtitles from Chinese short-form videos and rewrites them as Korean narration scripts. We wrapped an admin around it so operators can tune adaptation tone in real time without redeploying.
Bringing Chinese shorts into the Korean market was the most hand-on task in the entire workflow.
Transcribe subtitles, translate, adapt to Korean narration tone, repackage as SRT, then download the result and lay it over the video — all of it manual, every time. The more human time per clip, the less content the operator could push.
And tone wasn't a fixed target either. Each channel needed a different voice, so 'this adaptation came out great' was no guarantee for the next clip.
We collapsed the five manual steps into one automation pipeline, and pulled tone tuning out of code and into an admin prompt.
When a video uploads, PaddleOCR reads subtitles frame by frame and GPT-4o handles translation and adaptation. The results come out as three simultaneous formats — SRT, Chinese+Korean script, and Korean-only script — so operators pick what they need.
Changing tone no longer requires a deploy. Editing the prompt in the admin applies to the next clip immediately. Approved-only signups also let us track who processed which videos.
Tailwind-based responsive admin. From video upload to result download, one screen.
OCR, translation and adaptation chained as serverless events. Container images pre-load the model to cut cold starts.
PaddleOCR reads subtitles per frame, GPT-4o translates and adapts. Tone controlled from the admin prompt.
Job, Frame, User and Config unified in a single-table design. Approved-signup uses JWT custom auth.