How Accurate Are Auto Captions? A Creator's Checklist

Auto captions are accurate enough to be a useful first draft and not accurate enough to post unchecked. No single percentage tells you which side your video falls on. Accuracy depends on your voice, your microphone, your room and the words you use, so the only reliable answer is to test your own captions before they go live.
That is what this checklist is for. If you post podcasts or short-form video from an iPhone or iPad, you will find below what makes captions go wrong, a pass/fail list to run before you post, a quick way to spot-check a draft, and how to fix mistakes without rebuilding your edit.
How accurate are auto captions? Why there is no single number
Auto captions have no fixed accuracy rate, because results change with the speech-recognition system and with the recording it is fed. A clean studio voice and a windy street clip can produce very different drafts from the same tool.
The standard measure is word error rate (WER). It counts words the system substituted, left out or inserted, so a lower score is better. It is a useful yardstick for comparing systems on the same audio, but a headline WER from a vendor tells you little about your own footage.
You will also see 99% quoted as the gold standard. Treat that as an industry benchmark, not a promise. The FCC's rules for TV captions ask for captions that are accurate, synchronised, complete and properly placed, and they do not set a blanket 99% requirement for every caption (FCC, 2024). A tool that advertises a big number has not proved your video will hit it.
Assume every auto caption draft needs a proofread, and spend more time on it when your audio is hard.
What makes auto captions go wrong (accents, noise, names, jargon, overlapping speakers)
Auto captions go wrong most often when the audio or vocabulary is unusual: accents and dialects, background noise, overlapping speakers, and unfamiliar names or specialist terms all raise the error rate.
Accents and dialects. Regional UK speech, from Glaswegian to Geordie to Scouse, can be misheard, and so can slang. If you speak the way you speak at home, expect to correct more.
Background noise. Traffic, café chatter, wind and music under your voice give the system less to work with.
Names and brands. Guest names, place names, product names and your own channel name are the words a general-purpose system is least likely to have seen.
Jargon and numbers. Niche terms, acronyms, prices and dates are common trouble spots.
Overlapping speakers. Two voices at once, normal in a podcast with banter, tend to produce garbled or merged lines.
The nasty part is that a wrong caption often reads fluently. A clean-looking transcript is not evidence of an accurate one.
The creator's checklist: what to check before you post
Run this pass/fail list on every draft. If any line fails, fix it before posting.
Names pass. Every guest, place and brand name is spelled correctly, including your own.
Numbers pass. Prices, dates, times, percentages and counts match what you said.
Jargon pass. Specialist terms and acronyms are right, not swapped for a similar-sounding common word.
Meaning-flip pass. Words like "not", "can't" and "won't" are present where you said them and absent where you did not. One missing "not" reverses your point.
Hard-audio pass. Any stretch with noise, music or two people talking has been read closely, not skimmed.
Start and end pass. The first line and the last line are right, because viewers read those first and remember them.
Line-break pass. Captions split at natural phrase breaks and fit on screen without covering faces or on-screen text.
Sync pass. Words appear when they are spoken, not noticeably early or late.
If you can tick all eight, post. If you tick fewer than six, the draft needs a proper read-through, not a skim.
How to spot-check a caption draft quickly
Play the video while reading the captions, and stop the moment you see a word that looks off. That single habit catches most errors, because your ears will notice a mismatch your eyes miss on a silent read.
Then go looking for the high-risk words on purpose. Scrub through and verify names, place and brand names, specialist terms, numbers, and any word that could reverse the meaning, such as "not" or "can't".
For a short clip, one full watch-and-read pass is enough. For a long podcast episode you will not have time to proof every minute, so be deliberate. Check the opening, check any section with a guest, check every number, and sample the noisiest stretches. A transcript view helps here, because you can search for a name and jump to every place it appears. If you want to see how transcripts work in practice, see how to get a transcript of a TikTok video.
Fixing errors: edit the transcript, not the timeline
The quickest way to fix a wrong caption is to correct the text in a transcript view, not to nudge caption blocks around a timeline. Typing a corrected word into a transcript is one action. Dragging, splitting and retiming caption blocks is several.
That is the idea behind transcript-based editing: the words are the interface. You read, spot the mistake, retype it, and the caption follows. The same view lets you delete a sentence or a filler word and have the video cut to match, which is why the approach suits talking videos. We cover the method in edit video by editing text on iPhone.
Timeline-first editors can do the job too, and plenty of creators are happy with them. The trade-off is effort. Every text fix may also mean adjusting timing or layout by hand. If you proof captions on every post, the extra taps add up across a week.
Burned-in captions vs subtitle files: why it changes how much checking you need
Burned-in captions need the most checking, because once they are baked into the video you cannot change them without re-exporting.
Burned-in captions are drawn onto the picture. Viewers cannot switch them off, and a typo is permanent in that file. Short-form platforms such as TikTok and Reels usually call for this style, so proof these hardest.
Subtitle files are a separate text track. Platforms that support them let you upload a corrected file later, and viewers can toggle them. A mistake is easier to repair, though it still reflects on you while it is live.
The more permanent the format, the stricter the checklist. Run all eight checks on burned-in video. For subtitle files, still run the names, numbers and meaning-flip checks, since those are the errors that mislead people. For a walkthrough of adding captions in the first place, see how to add captions to iPhone video.
Where Cutscript fits: transcript editing on iPhone and iPad (coming soon, join the waitlist)
Editing video by editing the transcript is one option for creators who want to proof captions in text rather than on a timeline. It is coming soon to the App Store and is not available to download yet, so the way to get in line is the waitlist at sail.care/cutscript.
Here is what it is built to do, going by what we know. You edit the transcript and the video follows. It cuts filler words automatically and adds captions. It runs on device.
It will not be the only choice. CapCut and other editors generate auto captions too, and if you already work comfortably in one, the checklist above applies unchanged. For a side-by-side on one task, see Cutscript vs CapCut for removing filler words. Whichever tool drafts your captions, you still do the proofing.
Proof every draft, and join the Cutscript waitlist if you edit on iPhone or iPad
Trust auto captions as a draft and never as a finished product. The errors that hurt you (a wrong name, a dropped "not", a mangled number) are exactly the ones that read smoothly. Run the eight-point checklist on every post, and run it hardest on burned-in video. If you would rather fix captions by editing text on your iPhone or iPad, join the Cutscript waitlist at sail.care/cutscript and be among the first to try it when it reaches the App Store.