Fixing auto-caption mistakes: names and numbers
By AutoCap · 31 October 2026
Automatic transcripts go wrong in predictable places: names, brands, English inside Indian-language speech, numbers and pauses. The order to check them in.
Why the mistakes gather in the same places
A speech recogniser writes down the words it judges most likely for the sounds it hears. Everyday words in clear speech usually come out right. Errors collect in words the model rarely met in training and in sounds that fit more than one word, so a transcript is quicker to check by hunting for those than by reading every word with the same suspicion.
What to look for
Most corrections fall into one of these groups:
- Names. Surnames, places, guests and your own channel name. An unfamiliar name can come back as ordinary words that sound like it, or split in two.
- Brands and products. Many are coined words whose spelling the sound doesn't give away, so even a correct hearing can be spelt wrong.
- English inside Indian-language sentences. Where a speaker switches, the English word may be misspelt, joined to its neighbour or written in the other script. The guide to choosing a caption script helps you settle on one convention.
- Numbers. Prices, dates, quantities and phone numbers. A wrong digit looks as tidy as a right one, and AutoCap's automatic emphasis favours words with digits, so a mistaken figure may well be the highlighted word.
- Words that sound alike, such as kaam (work) and kam (less), or saas (mother-in-law) and saans (breath). Quick speech blurs the longer vowel or the nasal sound, and in romanised captions one letter is all that separates them.
- Missed pauses. AutoCap starts a new caption after about 0.28 seconds of silence, or after a full stop, question mark, exclamation mark or danda. If the timings or the punctuation miss a break the speaker made, a caption runs on into the next sentence.
An order that saves time
I'd work in this order, because it stops a later fix from undoing an earlier one:
- Watch once without stopping. The playback speed button steps through 1×, 1.25×, 1.5× and 2×, and the captions keep pace, so a faster pass is possible when the speech is clear. Note the numbers of the captions that look wrong.
- Fix names and brands everywhere they appear. There is no find-and-replace, so settle each spelling first and scan the caption list for it.
- Check numbers against the audio, then read each caption for meaning to catch lookalikes and mixed-language slips.
- Delete filler and false starts. Removing a word leaves a silence, and one longer than about 0.28 seconds can split its caption in two.
- Fix line breaks once the words are final: split captions where the speaker paused and move stray words to the line they belong on.
- Last, check the highlighted words. When a clip is transcribed, AutoCap marks likely emphasis words, and on a template with a spotlight it may give one word the whole frame. Either pick can be wrong, and the word's own buttons remove it.
Making corrections in AutoCap
Corrections happen in the Captions panel, which lists every caption with its start and end time. Double-click a word to retype it, or select it and press Edit text; Enter keeps the change and Escape abandons it. Retyping only changes the spelling. The word stays tied to the moment it was spoken, so a correction can't knock the timing out.
A selected word also offers Split line here, previous-line and next-line buttons, Delete word, and Emphasise or Spotlight where the template supports them. On a keyboard, Space plays and pauses, the left and right arrow keys step between words, Enter splits before the selected word, Backspace deletes it, and Cmd or Ctrl with Z undoes; add Shift to redo. The keys don't act while you are typing in a field.
One fix needs the timeline. If a name came back as two words, retype the first as the whole name, delete the second, then drag the end handle of the remaining block until it covers the full name; each block has a handle at both edges, and dragging one never moves its neighbours. To remove several captions at once, tick them in the caption list, Shift-clicking to select a run, and use the Delete button that appears.
What editing can't rescue
Editing replaces words the recogniser misheard; it can't recover speech it never caught. Under loud music, with people talking over each other, or with the microphone far from the speaker, whole passages can come back with the wrong words and the wrong number of them. Retyping those keeps timings that belong to words nobody said, so the highlighting still won't follow the voice.
Timing repair has a ceiling too. When the recogniser times only whole phrases, AutoCap estimates where each word falls and then shifts the boundaries onto quiet moments it can find in the audio. If background music is loud, or the voice is very soft or poorly recorded, it leaves the estimates as they were, and words after a breath it missed stay early. Why captions drift out of sync goes further into timing.
AutoCap can't be given a list of names in advance, so recurring names need checking in every video. And if auto-detect got the language wrong, don't correct line by line: upload the clip again and pick the language under What language is spoken? Editing costs no transcription minutes, but a second transcription uses them again.
Questions
Does correcting a word change when it appears?
No. Retyping changes the text and leaves the word's start and end exactly where they were. To move a word, drag the edges of its block on the timeline.
Why did a caption split in two after I deleted a word?
The deleted word leaves a silence, and AutoCap starts a new caption after a pause of about 0.28 seconds. Undo the deletion, or move each stray word back with the previous-line button.
Do corrections use up transcription minutes?
No. Minutes are counted when a clip is transcribed; fixing words, timings and line breaks afterwards costs none.
Can AutoCap learn the names I use often?
No. Every clip is transcribed without a custom word list, so the same names can go wrong each time. Keeping a note of their exact spellings makes the check quicker.