Lead story · Accuracy
How accurate are auto generated captions?
With one clear voice and a decent mic, most lines come out right the first time. The mistakes show up in the same places again and again.
Auto captions are written by a speech recognition model instead of a person. That model is the "AI" in any AI caption generator, and ours is Whisper large-v3 turbo. It listens to the audio, picks the most likely words, and marks when each one starts and ends. We then group the words into lines of up to about 42 characters and roughly 5 seconds. Vertical videos get shorter lines, around 22 characters, so they fit a phone screen.
So proofread before you export. On a short clip it takes a minute or two: click a line and the video jumps to that moment, so you can hear what was actually said. YouTube gives creators the same advice about its own automatic captions.