EnCaptions feel simple. The video plays, the words appear in sync, the viewer reads them. Anyone who has actually shipped captioned short-form video knows what's hiding behind in sync: a sequence of decisions about word grouping, display duration, font sizing, line breaking, and contrast — each of which is its own small disaster waiting to happen.
Here's how EnDroit's captioning pipeline handles them. Some of this is well-known; some of it is the kind of detail you only learn after you've watched 200 videos burn because the captions were unreadable on a metro screen.
Step 1: word-level timing
Most caption tools work at the phrase level. Whisper, MFA, ElevenLabs forced alignment — they all return per-word timestamps. Use them. Phrase-level captions can't keep up with a 3-words-per-second speaker, and the resulting "block of text appears, block of text disappears" rhythm is what makes amateur captioned videos feel amateur.
EnDroit gets word timings from one of three sources, in order of preference:
- ElevenLabs forced alignment, when you use the built-in voice synthesis
- Whisper (large-v3) word-level transcription, when you upload your own voice
- A linear estimate (length × 0.06 + 0.25s per word), only as a fallback when the audio file is silent or fails to transcribe
Step 2: word grouping (this is where it gets interesting)
Word-by-word captions are technically possible but unreadable. The eye can't keep up with single words appearing and disappearing at speech rate. The answer is to group words into reading chunks — typically 4 to 5 words per chunk, displayed for around 1.2 to 2.0 seconds.
The naive algorithm: every 4 words, cut a new chunk. This works 70% of the time and creates atrocious cuts the other 30% — words orphaned from their object, articles split from their noun, "ne pas" broken across two chunks.
What we ship in EnDroit is a slightly smarter version. The chunker has four rules, in order of priority:
- Break on punctuation. A period, comma, or "—" is a natural pause. Cut there if possible.
- Break on long silences. If the gap between two words exceeds 220ms, that's almost always a phrase boundary. Cut there.
- Soft target: 4–5 words. Aim for chunks of this size when nothing else dictates the cut.
- Hard ceiling: 2.5 seconds. Never display a chunk for longer than this, even if it means cutting between bad words. The viewer's eye has stopped reading at that point regardless.
Rule 4 is the one nobody talks about. People obsess about cutting in the right place; they forget that a perfectly-cut 4-second chunk is worse than a slightly-awkward 1.8-second chunk, because nobody reads a 4-second caption.
Step 3: keyword emphasis
For legal content specifically, certain words deserve visual weight. Article numbers, monetary amounts, jurisdiction names, and the words that carry the legal claim ("INTERDIT", "OBLIGATOIRE", "1 AN", "15 000 €") should pop visually so they're remembered after the video ends.
EnDroit's caption renderer scans each chunk for keywords from the script's "highlights" list. If a word matches, it's rendered at 76px instead of the default 68px, with an accent-colored glow. The viewer's eye latches on, the message lands, the audience retention curve tells you it worked.
Step 4: positioning and contrast
Captions for vertical video have one hostile neighbor: the platform's UI. The TikTok caption text and the description live in the bottom 280 pixels of the screen. Instagram Reels has a similar dead zone. Anything you display in that area is at best partially covered, at worst completely hidden.
EnDroit renders captions in the upper-middle third of the frame (the "safe band" we calibrated by overlay-testing on real TikTok). Every caption has:
- A semi-transparent black backdrop (rgba(0, 0, 0, 0.7)) for readability over any background
- A 2px white text-shadow ring to maintain contrast even on bright backgrounds
- A maximum line length of 26 characters before wrapping, tuned for 6-inch phone screens at arm's length
Step 5: the part where we throw most of it away
After all of this, the renderer outputs a sequence of caption frames. Then we audit them. If any chunk:
- Has a display duration above 3 seconds
- Has more than 7 words
- Was cut mid-phrase against rule 1
… we re-run the chunker with a tighter target. The audit catches the cases where the alignment data was noisy enough that the chunker made a poor call. About 1 in 8 videos triggers a re-run. We never ship a video where the audit still fails after two retries — we surface the error to the creator instead.
Why this much engineering for "captions"
Because retention on captioned short-form video is dominated by caption quality. We measured it on our beta cohort: videos with clean, on-time, properly-grouped captions average 67% retention. Videos with the same content but lazy captions (block-of-text or out-of-sync) average 41%. A 26-point retention delta is the difference between a video that gets recommended and a video that dies in your feed.
The viewer doesn't notice good captions. That's the goal. They notice bad ones, instantly, and they swipe.