Your transcript is nearly perfect, except the one word that matters most. Your name comes back as a common word that sounds like it. Your product becomes two random words. A guest’s surname turns into something faintly rude. Ordinary speech is the easy part for auto captions. Names, brands and jargon are where nearly every embarrassing error lives, and they’re also the words viewers are most likely to notice.
The short answer. Speech recognition swaps rare words for common ones that sound alike, so names, brands and jargon come back wrong far more often than everyday words. Keep a short list of the names you use, read the transcript once looking only for them, and correct each one before you export. Once captions are burned in, a misspelt name is in every copy you post.
Why names come back wrong
A transcription model picks the string of words that most likely matches the sound, weighted by how often it has seen those words before. It has heard “Katie” a huge number of times and your brand perhaps never. When the audio is close enough, the familiar word wins. The post on why auto captions get words wrong covers the general mechanism. This one is about the fix.
The useful thing to take from it: the mistakes are predictable. The model doesn’t produce gibberish. It produces a real, plausible word, which is exactly why these errors slip past a quick read.
The five ways names go wrong
Knowing the patterns makes them much faster to spot.
- The sound-alike swap. “Kiera” becomes “Kira”, “Aimee” becomes “Amy”, “Notion” becomes “motion”. The result is a real word or name, spelt correctly, and wrong.
- The split. Compound brand names get broken into ordinary words. “Shopify” can come back as “shop if I”, and “YouTuber” as “you tuber”.
- The wrong styling. The letters are right but the form isn’t: “Iphone”, “tik tok”, “Youtube”. Stylised names with odd capitals, numbers or no spaces almost never survive as written.
- The spelled-out acronym. You say “S-E-O” and get “SEO”, or you get “S E O”, or “seo”. Acronyms you say as a word, like “NASA”, tend to fare better than ones you spell letter by letter.
- The handle or number. Usernames, discount codes, model numbers and prices get normalised into whatever the model thinks is a sensible way to write them. A handle like “@maya.makes” will not come back as you typed it.
Before you record
You can’t fully record your way out of this, but you can make it easier.
Keep a spelling list
Write down every name you say often: your own, your brand, your products, regular guests, places, recurring jargon. Next to each one, put exactly how it should appear on screen, including capitals. For a creator this list is often only a dozen lines. It turns every future check from “does this look right?” into a straight comparison.
Say the hard ones clearly
Slow down slightly on names and leave a tiny gap either side. This helps most with names that sound like everyday words, because it stops the model running them into the next word. It won’t teach the model a name it has never seen, but it gives the word a clean start and end. That matters later, because the timestamp on that word decides when your corrected spelling appears.
Keep music off the name
If a track or sound effect lands on the exact moment you say your brand, the model has even less to go on. Add music after transcription where you can.
Some caption and transcription tools let you supply a list of expected words before you transcribe. If yours does, put your spelling list there. If it doesn’t, the list still does its job at the proofreading stage.
Catching every one
A normal read-through misses these errors, because the wrong words make sense in the sentence. A better method is a dedicated pass.
- Read the transcript once for names only. Ignore grammar and filler. Stop at every capitalised word and every place you know you said a name, and compare it to your list.
- Watch for words that are a bit too ordinary. If a sentence suddenly mentions “motion” or “shop if I” in a spot where you remember saying a product, that’s your error.
- Play the clip at the name points. Thirty seconds of listening settles most doubts faster than squinting at text.
- Check the first two lines twice. The hook is what everyone who stops scrolling reads, and a misspelt name there costs the most.
If your style uses all caps, casing errors disappear on screen (“TIKTOK” is fine however it was transcribed), but spelling errors are as visible as ever.
Correcting without breaking the timing
Fixing a word is usually a tap and a retype. In Capster you tap any caption to rewrite it, split the line, or adjust when it appears, and the step-by-step guide to adding captions on iPhone shows where that sits in the flow. A few things are worth checking after any correction.
When the fix changes the word count
“Shop if I” is three words; “Shopify” is one. When you merge a split brand name back together, make sure the corrected word still covers the whole time the name was spoken. Otherwise it flashes up for half the duration. The reverse happens too, when a name the model heard as one word is really two.
When the fix makes a line longer
A long brand name or a full name with a title can push a line past what fits on a phone. Split the line where it reads naturally, ideally with the name kept whole on one line.
When the style shows one word at a time
Word-by-word styles give each word its own moment on screen. A long name like “Wojciechowski” gets the same brief slot as “the”, which can be too fast to read. Check those words in the preview, and nudge the timing or pick a different layout from the caption styles if a name keeps blinking past.
A quick check before export
| Check | What to look for |
|---|---|
| Your own name and brand | Spelling and capitals exactly as on your list |
| Guests and places | Ask guests how they spell their name before you post |
| Products and jargon | Splits into ordinary words, wrong capitals |
| Handles, codes, prices | Symbols, dots and digits exactly right |
| Corrected words | Still on screen for the whole time they’re spoken |
| The hook | Read it twice |
This matters more for burned-in captions than for any other kind. A sidecar subtitle file can be fixed and re-uploaded. Once you burn subtitles into the video, the text is part of the picture, and the only fix for a wrong name is a new export and a new post.