01 / KNOW THE SOURCE
Start by finding out which kind of text you have.
A result carries a short note above the words. It names the platform, sometimes the language, and then the source of the text: usually Caption track when the page read a track the video already had, or Speech recognition when it listened to the audio. For a file you uploaded, the note says Your file and Speech recognition. The note ends with a sentence worth taking literally: check important words against the original video, or against the recording when the text came from a file.
The two sources go wrong in different ways. A caption track is text somebody wrote or generated to fit the picture, so it can be abbreviated or loosely timed. Speech recognition writes down what it heard, so it can mishear. Neither is a finished transcript, and the note is there to remind you of that.
02 / WHERE ERRORS HIDE
Five places where words go wrong.
- Names. Brands, people and places are not ordinary words, so they can be replaced by something that sounds similar.
- Numbers. Fifteen and fifty sound close, and a date or a price matters most exactly when it is wrong.
- Homophones. Their and there, to and two, may come out as the wrong word.
- Jargon. Technical terms may be swapped for common words that sound alike.
- Overlapping speech. When two people talk at once, one of them tends to drop out of the text.
Caption tracks add problems of their own. Captions that were generated automatically may contain errors nobody has corrected, and captions written by hand may shorten the speech to fit the screen. Either way, check the lines you need against the recording.
Improving the recording before you upload it is covered in the article on getting a more accurate transcript. This one is about what to do after the words come back.
03 / THE FIVE-MINUTE CHECK
A routine that fits in five minutes.
Put the video and the text side by side.
Play the video at normal speed with the text open next to it. If timings are available, leave Show timestamps ticked in the reader so each piece has a time beside it, which tells you roughly where in the video to listen. If the checkbox is disabled, replay the recording and find the passage by its words.
Stop only where it matters.
Pause at names, at numbers and at any sentence you plan to quote. You are not proofreading the whole video; you are checking the lines you will rely on.
Fix those lines in your own document.
WatchToText does not edit the words or move the timings, so corrections happen in the document you paste the text into. Add punctuation wherever the meaning could be read two ways.
Search for the same slip again.
If a brand name came out wrong once, it may have come out wrong elsewhere too. Search your document for the wrong spelling and fix them all before you move on.
Leave the rest alone.
Unless a line changes the meaning, it is not worth fixing. Time spent on harmless slips is time not spent on the names and numbers.
04 / RE-RUN OR EDIT
When to find a better copy instead of editing.
If one name is wrong, edit it. If a whole stretch is garbled, editing may not be enough, because the problem is usually the audio: music under the voice, wind, a microphone far from the speaker. The way to a better result is a better input, so look for a cleaner copy of the video.
If you have the original file, the upload route described in MP4 to text lets speech recognition work from that file directly, up to 20 minutes and 200 MB. The result still depends on how clear the audio in that file is.
05 / WHAT A CHECK CANNOT CATCH
Matching the words is not the same as checking the facts.
A check against the video catches words that were heard wrongly. It does not tell you whether the speaker was right. A transcript records what was said, not whether it is true, so a number that appears in the video exactly as the transcript has it can still be wrong.
If you plan to repeat a claim as fact, look for the source behind it. Quote the transcript to show what the creator said, and use a primary source to show what is correct.
06 / QUOTE CAREFULLY
Quote what you checked, and say where it came from.
Quote only the lines you have checked against the video. Name the creator and the platform, keep the quote short, and mark any cut with an ellipsis instead of quietly joining two lines.
Spoken language is loose: a transcript will contain repeats, half sentences and filler words. Either quote the words exactly as they were said, or paraphrase them in your own words, and say which one you have done. Mixing the two without saying so is how a quote ends up putting words in someone's mouth. If you quote in writing, link to the original video so readers can hear it for themselves.
07 / QUESTIONS
How do I know whether the text came from a caption track or from speech recognition?
The note usually says Caption track or Speech recognition. For a pasted YouTube transcript, it says Pasted from YouTube's transcript panel; a result without a more specific source label may say Transcript. For a file you uploaded it says Your file and Speech recognition.
Does WatchToText correct the mistakes for me?
No. It returns the text as it was read or recognised. It does not edit the words, move the timings or add captions to the video.
Is a check still needed when the video has captions?
Yes. The note above the words asks for the check on both kinds of text, and names and numbers are the first things to look at.
What if two people talk at once?
One voice often drops out of the text. Check that stretch against the video and add the missing words by hand.
READY WHEN YOU ARE
Read it once more, then quote it.
Get the transcript, check the names and numbers against the video, and keep the quote short.
Get a transcript ↗