Captions are the part of a short people read most closely, because most of them are watching with the sound off. That also makes captions the part where a mistake costs the most. A wrong word in a voiceover goes by in a fraction of a second. A wrong word in big white letters sits in the middle of the screen while the viewer reads it.
Most captions today start as automatic transcription, and that includes mine. Transcription has become very good at ordinary speech. It’s still weak at exactly the words your audience cares about most: names, slang, and anything said in an accent it wasn’t mostly trained on. This post covers where it goes wrong, how I check it, and the one rule I follow when I can’t be sure.
Why names are the first thing to break
A speech model doesn’t know who your guest is. When it hears a name it hasn’t seen often, it writes the most likely thing that sounds similar. Often that’s a common word, or a more common name.
I see this constantly. When I prepare sample clips for podcasts, I read the auto-captions of the episode first, and a misspelled guest or player name is close to the norm rather than the exception. An American football quarterback called Hurts came out as “Hurst” on one channel and “Herz” on another. On an anime fight-analysis channel I edit for, the character Buu came out as “Boo”, and a character called Recoome came out as “Raccoon”. In the comparison I wrote up in OpusClip vs. a human editor, one character’s name was misspelled three times in the tool’s captions for a single episode.
None of these are strange errors. They’re exactly what you’d expect from software that has heard “Boo” far more often than “Buu”. But the audience for a niche channel knows every name. A misspelled name in the captions tells them, in the first few seconds, that whoever made the clip doesn’t know the subject. It’s also the easiest thing to correct in the comments, so it’s often the first thing people do.
Slang and accents fail more quietly
Names are at least easy to spot once you look for them. Slang and accents are worse, because the wrong version often reads like a normal sentence.
One show I’ve cut clips for is a group of British hosts talking about American football. Their speech is full of London slang, and the transcription models I tried leaned American. A UK university sports league, BUCS, came out as “box” in every model I tried, because to an American-trained model that’s the obvious word. Short phrases like “not a ball knower” or “practice squad player” flipped between similar-sounding alternatives depending on which model I used, and sometimes from one run to the next.
The danger here is that the caption still makes sense. Nobody reading it stops and thinks “that’s a mistake”. The creator’s regular viewers do, though, because they know how the host actually talks, and a caption that sands down their slang into generic English makes the host sound like someone else.
Crosstalk is where models invent things
The hardest audio is several people talking over each other. That’s also where some of the best moments in a podcast happen.
On that same multi-host show, I ran two different transcription services over the same clips. Neither won outright. One of them inserted whole plausible lines in the crosstalk, phrases that fit the topic and sounded like something a host might say, but that nobody had said. The other got some of those spans right and invented different words elsewhere.
That’s the failure that worries me most. A misspelled name is a mistake. An invented sentence puts words in someone’s mouth, in large letters, on their own channel.
How I check captions now
My process isn’t clever. It’s mostly a matter of not trusting one source.
- Collect the names first. Before transcribing, I make a list of every name that’s likely to come up: hosts, guests, players, characters, places, the show’s own name. I take them from the video title, the description, the channel itself and the creator’s earlier captions. Some transcription tools let you pass in a list of expected words, and OpenAI’s own prompting guide for its Whisper model suggests giving it the correct spellings of unusual names for the same reason. It helps, but it doesn’t fix everything.
- Transcribe more than once. I use two different transcription models on the same audio. Where they agree, the words are usually right. Where they disagree, I know exactly which spots need a closer look.
- Settle every disagreement against the audio. For each disputed word, I check which version actually matches the sound, and then I listen to it myself in the context of the full sentence. Context matters: “threw” and “through” can sound identical, and only the sentence tells you which one makes sense.
- Check the finished short, not just the source. Once the clip is cut, I transcribe the exported short again and compare it with the captions. Cuts can clip the end of a word, and that changes what it sounds like. I once found a word-final “s” cut short enough that “face” was heard as “feet”. The fix was in the edit, not the caption.
- Read the captions on their own, with the sound off. That’s how most people see them. Anything that reads oddly gets one more listen.
When I can’t be sure, I leave it out
This is the rule I care about most. If a word is still disputed after all of that, I don’t guess.
A creator I work with on interview shorts records some conversations on a phone, and the audio can be rough. The rule I follow for those captions is simple: a word goes on screen only if the transcriptions agree on it, or I can confirm it by ear. If I can’t, it doesn’t get captioned, and if the whole line is uncertain, I cut around it or pick a different line.
The same goes for side comments in crosstalk. On the multi-host show, one host made a quiet aside I couldn’t confirm with any model or by listening again. It might have been a joke worth keeping. I left it out anyway, because a caption that might be wrong isn’t worth the risk of putting the wrong words in someone’s mouth.
Leaving something out sometimes costs a good line. Getting it wrong costs trust, and on a creator’s own channel it’s their trust, not mine.
Keep the creator’s voice, not a cleaned-up version
Getting words right isn’t the same as tidying them up. I caption what the person said, in the way they said it. If a host says “innit” or “washed”, that’s what goes on screen. I don’t swap slang for standard English, correct their grammar, or soften a phrase because it looks informal written down.
I do make a few small, deliberate edits. I drop filler like “um” and false starts that the cut has already removed from the audio, so the captions match what you hear. I write names the way the person or team writes them, not the way they sound. And when a creator has a habit in their own captions, like putting names in a different colour, I copy it rather than inventing my own. Choosing which words get that emphasis is its own job, and a topic for another post.
A short checklist
If you caption your own shorts, these are the checks that catch most of the problems:
- Every name is spelled the way its owner spells it. Check the title, description or a quick search, not your ear.
- Slang is written as it was said, not translated.
- Nothing in a crosstalk section is captioned unless you can actually hear it.
- The captions match the finished cut, not the original recording.
- Read once with the sound off before posting.
None of this is difficult. It just takes the time to treat captions as something you check rather than something the software hands you. It’s the same idea as picking moments carefully, which I wrote about in how to find short-form moments that work without the full episode: the defaults are a starting point, not the finished edit.
If you’d like to see your own names and slang captioned this way, I can make a free short from one of your videos. You keep it either way.