Fixing Auto-Generated Captions So They Read Like Real Sentences

The transcription engines got good. Whisper Large v3, Descript's 2026 model, the live tracks inside Premiere and DaVinci Resolve — they'll hand you a word-error rate low enough that you stop double-checking every noun. That's exactly why the captions still feel wrong on screen. The words are right. The reading is broken.

Captions aren't a transcript. A transcript is a record of what was said. A caption is a thing a person reads in one or two glances while also watching a face move and a b-roll cut land. Those are different jobs, and auto-caption tools only do the first one. Everything below is about closing that gap without re-timing the whole track by hand.

Why clean output still reads badly

Run any recent model on clean audio and you'll get accurate words split into caption events that ignore how sentences breathe. The three failures show up every single time:

• Line breaks that split a phrase. "I went to the / store and bought" — the break lands mid-thought, so the eye stalls.
• Events that are too long or too short. A nine-word caption flashes for 0.8 seconds; a two-word caption sits for four. Both fight the viewer's reading speed.
• Punctuation that's technically present but rhythmically wrong. The model drops commas where you'd pause and skips them where you actually breathe.

None of these are accuracy problems, which is why re-running the transcription never fixes them. You're editing the presentation layer, not the words.

Start with reading speed, not wording

Before you touch a single word, set a target. The convention most editors still work to in 2026 is somewhere around 15 to 17 characters per second for adult content, lower for kids' material. You don't have to memorize the number — every serious tool will flag events that exceed it. In Resolve, turn on the subtitle validation warnings. In a standalone editor like Subtitle Edit (still free, still the workhorse), the character-per-second column turns red when an event reads too fast.

Why start here? Because reading speed decides everything downstream. If a caption is too fast, the fix is usually to split it or extend it, not to reword. If it's too slow, you merge. Knowing the constraint keeps you from rewriting things that were fine.

The line-break pass

This is the single highest-impact edit, and it's the one auto-tools do worst. The rule I give every junior editor: break on grammar, not on width. Keep a full phrase together even if the top line ends up shorter than the bottom.

Good:

I went to the store
and bought way too much coffee

Bad, but what the tool gives you:

I went to the store and
bought way too much coffee

The eye wants to grab a complete unit — a noun phrase, a verb phrase, a prepositional phrase — in one movement. Break before a preposition ("and bought" stays together), before a conjunction, before the start of a new clause. Never split an article from its noun ("the / store") or a name across lines.

Two lines maximum per caption, always. A three-line caption is a paragraph, and nobody reads a paragraph off a face.

Let AI do the mechanical part, then govern it

Here's where 2026 tooling actually earns its keep. The models are genuinely good at reflowing punctuation and line breaks when you tell them the constraints. I run cleaned transcripts through a prompt like this before the manual pass:

You are formatting subtitle text for on-screen reading.
Rules:
- Max 2 lines per event, max 42 characters per line.
- Break lines on grammatical boundaries, never between
  an article and its noun or a preposition and its object.
- Add commas only where a natural spoken pause occurs.
- Do NOT change any words, add words, or fix grammar.
- Return the same number of events with the same timings.

Here are the events: [paste SRT block]

That last constraint is the load-bearing one. Left alone, a language model will "improve" the speaker — smooth a stumble, correct subject-verb agreement, turn spoken cadence into written prose. For captions that's a bug. The words on screen should match the words in the ear, or a deaf viewer gets a different video than a hearing one. Tell the model explicitly to preserve wording, then spot-check that it did. I've watched Claude and GPT both silently "fix" a deliberate fragment in dialogue; the instruction reins it in, verification catches the rest.

Handle disfluencies with intent, not a blanket rule

Ums, uhs, false starts, repeated words. The instinct is to strip them all. Resist it — decide per project. In a polished explainer, remove them; they add nothing and eat reading time. In a documentary interview or anything where character matters, a stammer can carry meaning, and cutting it flattens the person. What you should always remove is the transcription artifact — the "[inaudible]" tags, the doubled words the model hallucinated in a noisy passage, the stray "you know" it caught four times in a row.

Most editors have a filler-word removal toggle now. Use it as a first pass on clean-explainer content and turn it off for interviews. Don't let a global setting make an editorial decision you haven't made.

The timing touch-up nobody teaches

Two timing habits separate captions that feel professional from ones that feel auto-generated:

1. Respect shot changes. A caption that starts two frames before a hard cut and ends two frames after reads as a flicker. If an event's boundary lands within a few frames of a cut, snap it to the cut. Resolve and Premiere both expose shot-change markers you can snap subtitle edges to.
2. Give the last event a beat. Auto-tools end the final caption the instant the audio stops. Extend it 200 to 500 milliseconds so the viewer finishes reading after the speaker finishes talking. It's a small thing that makes the whole track feel considered.

A workflow that takes minutes, not hours

For a five-minute talking-head piece, this is the order I actually work in:

1. Transcribe with the best model available in your editor. Don't shop around — they're all close now.
2. Do a fast wording pass for proper nouns, product names, and any term the model can't know (client names, invented words). This is the only place accuracy still needs you.
3. Run the formatting prompt above to reflow breaks and punctuation.
4. Turn on the characters-per-second warnings and fix every red event by splitting, merging, or extending — not rewording.
5. Snap boundaries to shot changes, extend the final event, spot-check three or four spots against the audio.

Fifteen minutes for content that used to take an hour, and the result reads like someone cared. That's the whole game: the machine handles the words, you handle the reading experience, and you never let the machine make the editorial calls it isn't equipped to make.