Guide to AI-Generated Subtitles for Social Video

85% of Facebook video is watched without sound, and on TikTok more than 80% of users scroll with the sound off. Instagram autoplays muted.

Captions decide whether your first line lands or gets thumbed past.

#Subtitled video wins on the metrics that decide reach

The engagement case for captions is older than short-form video itself, and it keeps holding up. Subtitled social videos see roughly 40% higher watch time and about 16% more reach, and 80% of viewers say they're more likely to finish a video that has them. Add the 400 million-plus people worldwide with disabling hearing loss and the accessibility argument stands on its own, before any algorithm is involved.

The mechanism is simple enough. A muted video with no text on screen has one second to communicate through picture alone, which almost nothing does. A muted video with a caption has a sentence. Platforms measure the difference as retention, and retention is what feeds distribution, so the caption track ends up being a ranking input dressed as an accessibility feature.

#Speech recognition and script-based captions fail in different places

Two mechanisms produce automatic captions, and they break differently.

Speech recognition listens to finished audio and writes down what it hears, which puts every proper noun at the mercy of the acoustics. Accuracy runs 90-95% on clear speech and drops with an accent, background music, or a product name the model has never encountered. Script-based captioning starts from the text that generated the voice, so it cannot misspell anything the script spelled correctly.

Orange takes the second route, because the script exists before the audio does. AI subtitle generation under that arrangement is a timing problem rather than a transcription problem. The distinction shows up hardest on brand names and numbers, which are the two things a viewer notices immediately when they're wrong and the two things a transcription model guesses at worst.

#Word timing comes from the voice, and the words come from the script

The premium voice engine hands back a timestamp per word alongside the audio. The standard engine returns audio alone, so Orange runs a separate timing pass over that audio purely to time words it already knows. The transcript never becomes the caption text.

Sometimes the voice and the script disagree on how many tokens a phrase is, because the voice read "24,700" as five spoken words. Orange reconciles the two so that one awkward token degrades on its own instead of dropping the whole scene onto guessed timing, and every render records which timing path ran. A drifting caption has a cause you can look up rather than a mystery to shrug at.

#Burned-in captions are the only ones that survive a repost

Burned-in means the caption is rendered into the pixels, so it travels with the file. Download a Reel and repost it as a TikTok video and the platform caption stays behind while a burned-in one arrives intact. Sidecar SRT files are worth uploading on YouTube anyway, since the text gets indexed, but treat them as a supplement rather than the plan.

#Caption size is a share of frame height, not a fixed point size

Orange sizes vertical captions as a share of frame height rather than a fixed point size. Render the same job at a smaller resolution and the font size, margins, outline, and shadow all rescale, with the characters-per-line limit recomputed from the real usable width, so a smaller export doesn't ship text running off both edges.

The one-word style clamps further. It shows a single word at a time and cannot wrap, so the font shrinks until the longest word in the narration fits. Position is bottom-center, held above the per-platform UI band that Orange reserves per aspect ratio, which is the strip where the TikTok caption box and the Reels action rail sit. Captions placed by eye in a desktop editor land underneath that furniture surprisingly often.

#Non-Latin scripts need the font embedded or the export ships boxes

A subtitle renderer with no glyph for a character draws a tofu box, and it does this silently. Orange carries the right fonts for CJK, Arabic, and Devanagari scripts through the render itself, so a Japanese or Hindi video ships its own typography rather than depending on whatever the encoder happens to have.

Czech and Slovak get a second correction that matters for anyone posting in 34 languages. Numbers are spelled out phonetically in the text sent to the voice engine, because that's what makes a voice pronounce them naturally. The caption shows the written form instead. The screen reads "24 700 kWh" while the audio says it the way a Czech speaker would, rather than the caption reading back a wall of spelled-out digits.

#The highlight should stop with the voice, not with the scene

Word-by-word highlighting has a failure mode nobody notices until they look for it. The last word of a line keeps lighting up while the visual holds for another second, because the caption's end time was stretched to fill the scene. Orange derives the final word's fill from measured speech end and holds the completed line until the cut, so the highlight tracks the voice while the text stays readable through the outro pad.

Five styles ship today: bold, minimal, animated, one-word, and boxed. For short-form, the animated word-by-word fill earns its keep, because the moving highlight gives a muted viewer something to track. Longer educational cuts read better with full sentences. Auto-subtitles run on every render regardless of style, with no extra step to trigger.

#A broken caption file should cost you captions, not the video

A structurally broken caption file used to be able to take the entire render down with it. Two layers now stand in the way. A structural pre-check drops an unparseable file before the burn and renders caption-free, and a burn that still fails triggers one re-render with subtitles switched off.

A video without styled captions beats a failed job and a spent credit balance. Both degrade paths get recorded on the job so the cause is visible afterward instead of showing up as a mysteriously plain video.

#Best practices worth keeping

1. Put captions on everything. Every social video, no exceptions.

2. Use high-contrast text. White with a dark outline or a background bar survives any visual underneath. Thin light-colored type does not.

3. Cap it at two lines. More than that and a phone viewer reads instead of watching.

4. Time it tightly. Words should appear as they're spoken. Even 200ms of lag reads as broken.

5. Check the bottom edge. Platform UI eats the lowest strip of the frame, and the amount varies by platform and aspect ratio.