On this page
Automatic transcription gets about nineteen words in twenty right, which is good enough to look finished and not good enough to publish. The missing word is almost never random: it is a brand name, a number separated from its unit, or a sentence cut in the wrong place.
Those are the words carrying your credibility. Which is why "the captions are automatic" quietly means "the one place I was specific is the one place it is wrong".
What are the four error classes?
Every project produces the same four. From this one's own caption files:
| Class | What it looks like |
|---|---|
| Brand and product names | a tool name rendered as a plausible English word, or respelled |
| Numbers and units | "95 %" split so the percent sign ends up on its own line |
| Language detection | an English clip transcribed as another language entirely |
| Sentence boundaries | a caption ending mid-thought, so the line reads as nonsense |
The first is the one that matters and the one no amount of spending fixes. A model has no way to know that a product is spelled the way you spell it, and it will substitute the nearest real word with confidence every time.
The third surprises people. A short clip, an accent, or a few seconds of silence at the start is enough for the detector to pick the wrong language, and the output is not garbled — it is a fluent, confident transcription of something you did not say.
What does it actually cost in time?
Fifteen to twenty minutes for a five-minute video, the first time.
But the time does not scale with length the way you expect. It scales with density of names and numbers. A five-minute script naming nine tools and six figures takes longer than a ten-minute one that names none.
| Video | Realistic check time |
|---|---|
| 5 min, conversational, no names | ~5 min |
| 5 min, naming tools and prices | ~20 min |
| 29 videos of the second kind | ~10 hours |
Ten hours is a real number, and it is the one that makes people skip the step.
The whole list
67 tools, what each costs and when you actually need it. Updated twice a year.
How do you make it five minutes instead of twenty?
A correction list, applied before you read anything.
Every brand name, product, and recurring figure you use goes into a find-and-replace file once. Run it over the transcript, then read. You are now checking rather than proofreading, and the difference in time is roughly four to one.
This works with every tool, costs nothing, and improves on its own — each video adds whatever new word it got wrong. It is the single highest-return habit in this whole pipeline, and it takes one file.
The remaining manual pass is for the two classes a list cannot fix: sentence boundaries, and the occasional clip the detector got wrong in its entirety. Both are obvious at a glance once the names are right.
Does paying more fix it?
Partly, and not the part you care about.
A better transcription tool improves general accuracy, handles accents better, and splits sentences more sensibly. None of them reliably knows how your tools are spelled, because your tools are not in the training data in the proportion your script uses them.
So the money buys you fewer errors in class two and four, and roughly the same number in class one. Given class one is where the damage is — a tool name misspelled in a video recommending that tool is a specific kind of embarrassing — the correction list is the better investment than the upgrade.
What an hour of transcription actually costs, compared covers the tools themselves when the transcript is the deliverable rather than the subtitle.
Is it worth doing at all?
Yes, and the reason is not accessibility — though that is reason enough on its own.
A large share of short-form video is watched with sound off. For those viewers the captions are not a subtitle track, they are the script. An error in them is not a typo in a supporting file, it is a mistake in the only version of the words most people will receive.
Put differently: if you would not publish an article with the product names misspelled, the bar for captions is the same bar.
What about burning them in?
Do it for short-form, skip it for long-form.
Burned-in captions survive every platform and every muted autoplay, which is what short-form needs. On long-form they fight the platform's own caption track, cannot be turned off, and prevent the automatic translation that gets you viewers in other languages.
The practical rule on this project: vertical clips get captions rendered into the frame, horizontal videos get a caption file uploaded alongside.
Where this came from
The error classes are from this project's own caption files, produced with a small local transcription model across 29 long-form videos and 13 short clips in October 2026, and corrected by hand. The time estimates are my own and are honest rather than flattering.
Can you edit video without a monthly subscription? and how do you make thumbnails without a designer? cover the rest of the pipeline.
The AI Persona Playbook is the one about running the channel this feeds.
Tools in this piece
| Tool | What it does | Price |
|---|---|---|
| Notta | Transcribes meetings, calls and interviews, then pulls the decisions out of them. | Free tier, then ~$8/mo (yearly) |
| Transkriptor | Transcribes meetings, interviews and calls, with summaries and search on top. | From ~$9.99/mo |
| HappyScribe | Transcribes, subtitles and translates video with AI, with human linguists as an option. | €0.20/min, or from €17/mo |
| Descript | Edit video by editing the transcript. | Free tier, then ~$15–30/mo |
| CapCut | Fast shorts and vertical editing, with captions. | Free, Pro ~€10/mo |
Every tool has its own page with the price, who should skip it and what to check before paying. Some links are affiliate links.
Questions people ask
Do you still need to check automatic captions?
Yes. Automatic transcription is close enough to look finished and wrong in predictable places — brand names, numbers with units, and sentence boundaries. The errors cluster exactly where your credibility is, which is why skipping the check is expensive.
What does checking captions cost per video?
Roughly fifteen to twenty minutes for a five-minute video, most of it spent on proper nouns. The time does not scale with length as much as with how many names and figures the script contains.
What kinds of errors does automatic transcription make?
Four classes: brand and product names it has never seen, numbers separated from their unit, the wrong language detected on a short or accented clip, and sentences split in the wrong place so a caption ends mid-thought.
Does a more expensive transcription tool fix this?
Not the brand-name class, which is the one that matters most. No model reliably knows how your tools are spelled. A glossary or find-and-replace list fixes it in seconds and works with any tool.
Are captions worth the time at all?
Yes. A large share of social video is watched muted, so unchecked captions are not a subtitle problem, they are the script as most people will receive it.
What is the fastest way to speed this up?
A correction list. Every brand name, product and recurring figure in a find-and-replace file, applied before you read anything. It turns twenty minutes of reading into five minutes of checking.