Suppose you hear a line like this in a video.

We meet on Friday. Sorry, Saturday.

The Japanese subtitles on screen read “We'll meet on Saturday.” Did your ears, which heard Friday, get it wrong, or did the subtitles? It's tempting to suspect one or the other. But neither is wrong.

The speaker really did say Friday, then immediately corrected themselves to Saturday. As an utterance, Friday definitely exists, and as a plan, Saturday is correct. The subtitles are simply translating the conclusion after the correction.

So when audio and text disagree, the question is not “Which one is right?” It's “What do I want to know?” The text you can trust changes depending on that question.

  • Content: In the end, when are you meeting?
  • Utterance: What was actually said?
  • Timing: When were those words spoken?

Subtitles and transcripts come in different kinds, and each is good at answering a different question. Misjudge that, and your check will be off even when you're looking at the right text.

Translated subtitles translate the conclusion

Japanese subtitles are a translation of the English. Within limited characters and display time, they drop hesitations and corrections and sum up the content that ultimately comes across. Translating “Friday — sorry, Saturday” as “We'll meet on Saturday” is accurate as content.

It isn't suited, though, to checking what you heard. The Friday that never appears in the subtitles can look as if you simply misheard it. The subtitles aren't bad; they're just doing a different job.

English transcripts are neatly polished, too

Just because something is written in English doesn't mean it's exactly as spoken. A transcript that prioritizes readability, drops the Sorry and the correction, and writes only We meet on Saturday. is perfectly ordinary.

That it's an official transcript doesn't mean it's word for word, either. For example, TED's official help explains how to find interactive transcripts on TED.com that are linked to the playback position in a video. It's a useful feature, but that explanation alone doesn't mean the timestamps or every utterance will match a podcast version of the same name. If the version differs, the corresponding text differs too.

In the end, was that text made to preserve the words as spoken, or to convey the content in an easy-to-read way? That is the dividing line.

Auto-captions are the result of speech recognition

YouTube's auto-captions are created using speech recognition based on machine learning. YouTube Help notes that they may not be displayed correctly depending on pronunciation, accent, dialect, background noise, and other factors. From the start, they are not meant to be treated as a perfect transcript.

If a correction remains, in a form like Friday sorry Saturday, it can be a clue that a correction happened. But corrections and hesitations don't always remain, and Friday itself may be displayed as a different word. The only way is to use them as text in which correct and incorrect parts are mixed together, checking them against the audio.

Sometimes the script came first, or the version is different

What was a Friday plan in a script written before recording may have been corrected during the actual recording. If there are versions with parts edited out, or versions in a different form, such as a video and a podcast, which recording your transcript corresponds to also matters. When you find a transcript, check the page's description to see which version it corresponds to. Skip that, and your checks will keep being off.

Subtitles can also appear early

Subtitles are displayed together, one sentence at a time. If “We'll meet on Saturday” appears early, while the speaker is still in the middle of saying Friday, it can look as if Friday was translated as Saturday. In this case, the content matches, and only the display timing is off. Listen again to a few seconds before and after, and see whether a correction follows, and you'll know.

How to choose is simple. If you want to know a plan or the conclusion of the talk, translated subtitles are often enough. At a spot with a correction, though, listen to what comes before and after and check that the conclusion matches the subtitles. If you want to check what you heard, use English text written as spoken, and check as far as whether that transcript is written in a way that preserves corrections. If you want to know when a line was said, don't rely on the subtitles' display time; move the playback position and check with the audio.

Friday definitely exists in the utterance, and Saturday is the conclusion of the plan. You don't need to make one the answer and discard the other. If you look at whether the text in front of you is meant to preserve the utterance or convey the content, a discrepancy becomes a clue rather than confusion. For choosing materials, including whether a transcript is available, five perspectives for looking at published practice materials covers the topic in detail.