Correcting auto-captions: names, numbers and more
By Captionwale · 31 October 2026
Auto transcripts slip in the same spots: names, brands, English words in Indian-language speech, numbers and missed pauses. What to check, and in which order.
Where errors cluster, and why
Speech recognition works by guessing the likeliest words for the sounds it receives. Clear speech made of common words mostly comes out fine. The slips gather around words the model met only rarely while it was trained, and around sounds that could be heard as more than one word. Hunting for those spots is quicker than reading the whole transcript with the same level of doubt.
The usual suspects
Nearly every correction belongs to one of these types:
- Names: of people, of places, of guests, and of your own channel. When the model doesn't know one, it may hand back ordinary words with a similar sound, or break the name in half.
- Brands and products. Lots of them are made-up words, and you can't tell how they are spelt by listening, so even a correct hearing may be spelt wrongly.
- English inside an Indian-language sentence. Right where the speaker changes language, the English word can come out misspelt, glued to the next word, or in the wrong script. The script-choice guide helps you fix one rule for the whole video.
- Figures, such as phone numbers, prices, quantities and dates. An incorrect digit looks just as tidy on screen as the right one. On top of that, Captionwale's automatic emphasis leans towards words with digits in them, so a wrong number could well be the one in the highlight.
- Near-identical words, for example saans (breath) and saas (mother-in-law), or kam (less) and kaam (work). Rapid speech smudges the nasal sound or the longer vowel, and in romanised text the gap between them is one letter.
- Pauses that were missed. A new caption starts when the speaker goes quiet for about 0.28 seconds, and after any ., ?, ! or danda. When a break the speaker made shows up in neither the timings nor the punctuation, the caption runs straight on into the following sentence.
Check in this order
Work through it in this sequence, so no later change undoes an earlier one:
- First, one uninterrupted viewing. The speed button has four settings, from 1× up through 1.25× and 1.5× to 2×, and the captions stay in step at each, so clear speech can go faster. Write down the numbers of any captions that seem wrong.
- Next, names and brands, everywhere. The editor can't find and replace, so decide each spelling up front, then run down the caption list looking for it.
- Then the figures, checked against the audio, followed by a read of each caption for sense, which catches near-identical words and language slips.
- After that, cut fillers and false starts. A deleted word leaves a silent gap, and if that gap tops 0.28 seconds or so, the caption may split in two.
- Once the wording is settled, sort out the breaks: cut a caption wherever the speaker took a pause, and shift any word sitting on the wrong line.
- Last, the highlighted words. At transcription time, Captionwale tags the words it thinks carry emphasis, and a spotlight template may hand one word the whole frame. Both guesses can miss, and the buttons on a word take its mark away.
How to edit in Captionwale
All correcting happens in the Captions panel, where every caption is listed with the time it starts and ends. To retype a word, double-click it, or pick it and use Edit text. Press Enter to keep the edit or Escape to drop it. Only the spelling changes; the word stays pinned to the instant it was spoken, so your fix cannot shift the timing.
A selected word also gives you Split line here, a button each for the previous and next line, Delete word, plus Emphasise or Spotlight on templates that have them. Keyboard shortcuts: Space toggles playback, and the arrow keys (left and right) hop from word to word. Backspace removes the selected word, Enter breaks the line just before it, Cmd+Z or Ctrl+Z undoes, and adding Shift redoes. Shortcuts are ignored while a text field has focus.
One repair needs the timeline. When a single name has come out as two words, type the full name into the first, delete the second, and stretch the surviving block by its end handle until it covers the whole name. Both edges of a block carry a handle, and pulling one never moves the blocks around it. To clear a batch of captions, tick each one in the list, Shift-click to take in a run, and hit the Delete button that then appears.
Problems editing can't solve
Editing swaps out words that were misheard. Speech that was never caught in the first place can't be recovered that way. Under loud music, with voices overlapping, or with the mic a long way from the speaker, entire passages can come back wrong in both wording and word count. Retyped, they still carry the timings of words nobody spoke, and the highlight won't track the voice.
Repairing timing has its ceiling too. If timings come back for whole phrases only, Captionwale estimates the position of each word and then snaps the boundaries to quiet moments it finds in the audio. Where the music underneath is loud, or the voice faint or badly recorded, it leaves the estimates untouched, and any words after a breath it failed to spot remain early. More on timing is in captions out of sync.
There is no way to preload names, so repeat names have to be checked video by video. If auto-detect picked the wrong language, don't correct the captions one by one; upload the clip afresh and choose the right one under What language is spoken? Corrections cost no minutes; a second transcription does.
Questions
If I correct a word, does its timing change?
No. Only the text changes; the start and end times of the word stay put. To shift it in time, pull its block's edges on the timeline.
Why did deleting a word split my caption in two?
The gap it leaves is silence, and silence of roughly 0.28 seconds makes Captionwale start a fresh caption. Undo the deletion, or use the previous-line button to pull each orphaned word back.
Will my corrections use transcription minutes?
No. Minutes are counted only when a clip is transcribed. Changing words, timings or line breaks afterwards uses none.
Can I teach Captionwale names I use a lot?
No. There is no custom vocabulary, so a name can go wrong in every clip. A note of the exact spellings makes each check faster.