Subtitles, Captions and Turkish-English Localisation

Turkish runs about twelve percent longer than the English it came from, and that is what breaks caption files. A line that fitted 42 characters in English arrives at 47, the renderer wraps it onto a third row, and that row sits underneath the TikTok caption bar where nobody reads it. The work here is timing and typesetting. Translation is the part that happens in the middle. The thresholds: Forty-two characters per line, two lines, reading speed capped at 17 characters per second for general audiences and lower for anything technical. Minimum event duration one second, minimum two frames between events, no event crossing a shot change unless the sentence genuinely does, shot-change snapping inside four frames. These are not house preferences. They are what broadcast and OTT QC rejects files on, and they happen to be the same thresholds that make a feed ad readable at arm's length on a phone. The Turkish problem nobody catches: Dotted and dotless i. İ and ı are separate letters, and a font missing those glyphs in the weight you actually use, or a text-transform to uppercase running under the wrong locale, turns İSTANBUL into ISTANBUL or the reverse. We check the font binaries for coverage of ı, İ, ğ, ş, ç, ö and ü across every weight in the deck, then prove it as a rendered sheet rather than a list of font names. Burned-in captions get the same check as pixels. Formats: SRT and WebVTT for platform upload. iTT and EBU-STL where a broadcaster or an OTT partner specifies them. Burned-in versions framed against current safe areas, because TikTok's interface takes the lower third and a strip down the right, and a caption centred in the frame is not centred in the visible frame. Localisation is not translation: An automotive spot going into a European market keeps WLTP figures and drops EPA ones, and the finance line is rewritten because the payment structure is genuinely different. Furniture goes to centimetres and to the local delivery promise, which is the line that actually sells it. A developer-tool video showing Cmd shortcuts to a Windows-majority audience shows Ctrl. Prices change currency and they also change the number, because a converted price reads as a converted price. Refused: Auto-captions tidied up and delivered as a subtitle file. Machine translation of ad copy, where register carries the message and a model returns something grammatical and dead. Synthetic dubbing and voice cloning. And languages nobody here can read: we manage linguists we have worked with for years in German and Arabic, and we tell you directly that our own QC on those is structural — timing, line length, glyph rendering — not editorial. Not our discipline: Transcription of archive footage as a standalone product. Sign language interpretation. Audio description for accessibility compliance, which is a separate craft and we will point you at people who do it properly. Re-editing picture to make room for captions; when a super collides with a caption we flag it, and either your editor moves the super or it becomes an edit job with its own quote. Poor fit: Anyone buying a language count rather than a language. Teams treating captions as the last checkbox before a Friday launch, since the second read happens at full speed against the finished picture and that takes the time it takes. Process: You send locked picture and the script if one exists. We transcribe against the actual audio rather than the script, time it, typeset it, and then a second reader watches it at full speed instead of scrubbing it in the editor. Delivery is the caption files, the burned-in renders per ratio, and the glyph proof.