Fixing Auto Generated Captions Before You Publish

Auto generated captions get you 90 percent of the way in seconds, and the last 10 percent is where viewers notice. A misheard product name, a missing speaker change or a caption that sits on screen for half a second is enough to make a deaf viewer give up on your video.

I run captioning for an archive that publishes several hours of footage a week, and nearly all of it starts as a machine draft. This is the editing pass we use to turn auto generated captions into something we are willing to put our name on, with realistic times for each step.

What auto generated captions get wrong most often

Speech recognition has improved a lot, and on a clean studio recording with one speaker the word accuracy can reach 95 to 98 percent. That sounds high until you do the maths. A 10 minute talk runs about 1,500 words. At 96 percent accuracy that is 60 wrong words, roughly one every ten seconds.

The errors are not random. In our logs they cluster in five places:

• Proper nouns. Names, brands, places and jargon. "Annodex" has come out as "and a decks" more times than I can count.
• Numbers. "Fifteen" and "fifty" get swapped, and years are often split into two words.
• Overlapping speech. When two people talk at once, the engine picks one or merges them into nonsense.
• Punctuation and sentence breaks. Long run-on captions with no full stops are hard to read even when every word is right.
• Non-speech sound. Laughter, music and a door slam are usually missing entirely.

A four pass method for fixing auto generated captions

Trying to fix everything in one watch-through is slow and you miss things. Splitting the work into passes, each with one job, is faster overall. For a 10 minute clip, budget 40 to 60 minutes total.

Pass one: find and replace the vocabulary

Before you press play, export the caption file as SRT or VTT and open it in a text editor. Search for every name and term you know appears in the video and fix them in bulk. Keep a running glossary per channel; ours has about 200 entries. This takes 5 minutes and removes a third of the errors.

Pass two: listen at normal speed and fix words

Play the video at 1x with the caption editor open. Pause on every error and correct the text only. Do not touch timing yet. Speeding up to 1.5x saves time on clean audio, but on interviews it causes you to miss small words like "not", which flip the meaning of a sentence.

Pass three: fix line breaks and timing

Now work through the file caption by caption. Each caption should hold one idea, two lines at most, and stay on screen long enough to read. Break lines at natural phrase boundaries, never between an article and its noun. Our rules on caption reading speed and line length give the exact numbers: around 42 characters per line and no more than 17 characters per second for general audiences.

Pass four: add speakers and sound

Add speaker labels wherever the speaker is off screen or a change would be unclear. Add sound descriptions in square brackets for anything that carries meaning: [phone buzzing], [audience laughs]. If you produce interviews or a video podcast, see our guide to labelling speakers in captions for a consistent format.

Tools that make the editing faster

You do not need paid software. YouTube Studio, the free Subtitle Edit on Windows and Aegisub on any platform all let you edit text and timing against the waveform. A few features are worth checking for:

Whatever tool you use, save the finished file in a format that keeps your edits when you re-upload or move platforms. Our comparison of caption file formats that survive re-uploads explains why SRT is a safe default and when VTT is worth it.

Why "just upload a transcript" is not the shortcut it looks like

A common tip is to skip editing the machine captions and instead paste your script into the platform and let it auto sync. It sounds efficient. In practice it often produces worse captions.

The reason is that a script is what someone planned to say, not what they said. Presenters ad lib, repeat themselves and skip lines. Auto sync tools force the script onto the audio anyway, so you get captions that are perfectly spelled and confidently wrong, sometimes drifting by several seconds after the first deviation. On one 20 minute webinar we tested, the synced script was out of step by 4 seconds by the halfway mark.

Script sync does work for tightly read material such as voiceovers and narrated explainers where the reader sticks to the page. For anything conversational, start from the machine draft and edit it. You will spend less time than you would fixing drift.

A quality check before you publish

When the four passes are done, do one final watch with the sound off. This is how many of your viewers will experience the video, and it exposes problems you cannot hear: captions that cover a lower third graphic, a joke that makes no sense without tone, or a speaker change you forgot to mark.

Finally, read the caption file once as plain text, without the video. Typos and missing words that slip past you while watching jump out on the page.

Spot check three random minutes against the audio. If you find more than two errors in those three minutes, something went wrong in pass two and it is worth another listen. For archive work we aim for 99 percent word accuracy, which on a 1,500 word video means 15 errors at most.

Frequently asked questions

How accurate are auto generated captions on YouTube?

On clear audio with one speaker, accuracy is often in the mid 90s by word count. It drops sharply with accents the model has heard less, background music, crosstalk and specialist vocabulary. Even at 95 percent, a 10 minute video can contain 70 or more wrong words.

Can I edit auto generated captions after a video is published?

Yes. Most platforms let you open the caption track, edit it and republish without changing the video itself. Download a copy of the edited file as well, so you are not locked in if you move or re-upload the video later.

How long does it take to fix auto captions?

Plan on four to six minutes of editing per minute of video for interviews and panel content. A single clear narrator can drop that to two or three minutes per minute. Building a glossary of names and terms speeds every later video on the same channel.