HOW VIMEO IMPLEMENTED AI-POWERED SUBTITLES

Summary from the TLDR Dev newsletter

When Vimeo built LLM-powered subtitle translation, they hit a counterintuitive bug: the AI would merge fragmented speech into one clean sentence, leaving empty time slots and blank screens mid-video. The fix was splitting the pipeline into three phases: chunk the source into logical thoughts, translate freely with no structural constraints, then use a separate LLM call just to remap the translated text back into the correct number of timed lines.