Saad Al-TamariDubljly turns an English lecture, from a YouTube link or an upload, into Arabic: a synced voice-over, Arabic and English subtitles, a study pack, and a tutor that knows the lecture. I built it solo, as one FastAPI service with a React front end and no extra infrastructure.
Why joining clips drifts
Arabic often takes longer to say than the English it translates. Join the voiced clips end to end and every sentence that runs a little long pushes the next one later. A few minutes in, the narrator is describing a slide that has already gone.
Place every clip on the original timeline
Each transcript segment keeps its original start time. Its Arabic clip is placed at that start, never after the previous clip. When a clip is longer than its slot, a fit ladder shortens or speeds it up until it fits, so the voice stays with the speaker for the whole lecture.
Learning starts before the dub finishes
The study pack (summary, key terms and their Arabic meanings) is built before anything is translated or voiced. The target is a study pack in about 30 to 40 seconds and a full dub in about 2 to 3 minutes for a 10-minute lecture. Those are spec targets, not measured figures.
Swappable providers, validated output
Speech-to-text, language models and voice each sit behind an adapter, so the pipeline never imports a vendor SDK. Structured calls ask for JSON and validate it with Pydantic: invalid output gets one retry with the error attached, then the router moves to the next provider.
The hard part of Arabic dubbing isn’t translation. It is time.