Why raw transcripts are hard to read
Captions are written to be shown a few words at a time, on screen, in sync with speech. That shapes the text: each line is short, many lines start with a time label, and the line breaks follow the timing of the audio, not the grammar. Read as a document, it feels choppy.
Cleaning is the step between "I have the transcript" and "I can read, quote or paste it into something". Done well, it changes layout only. It never rewrites what was said.
What the cleaner removes
- Sound cues. Square or round bracket labels for music, applause, laughter, cheering, snicker, sigh, gasp, screaming, sound and audio.
- Timestamps. Times written as
[00:00],(00:00)or0:00 -at the start of a line, with or without hours. - Extra whitespace. Repeated spaces and runs of blank lines.
- Broken lines. Lines are joined until a sentence ends and the paragraph has reached a sensible length, then a paragraph break is added.
Each of these is a separate option in the transcript cleaner, so you can keep the timestamps and still tidy the lines, or the reverse.
Worked example: before and after
The text below is original sample wording written for this guide, not a real video. It shows the same mess a real transcript has. The "after" output is produced by the same cleaning code the tool uses.
Before: raw lines
[00:00] Today we are looking at how to take notes from a long talk [Music] [00:05] without losing the parts that matter. (Applause) 0:11 - First, decide what question you want the talk to answer. [00:16] Then read for that answer [00:19] and ignore the rest until later.
After: all four options on
Today we are looking at how to take notes from a long talk without losing the parts that matter. First, decide what question you want the talk to answer. Then read for that answer and ignore the rest until later.
With only the timestamp option on, the cues and line breaks stay, which is useful when you want to keep a trace of where things were said:
Timestamps removed, nothing else
Today we are looking at how to take notes from a long talk [Music] without losing the parts that matter. (Applause) First, decide what question you want the talk to answer. Then read for that answer and ignore the rest until later.
What cleaning cannot do
- It does not correct mis-heard words, names or numbers.
- It does not add speaker names, because captions rarely include them.
- It does not translate, summarise or paraphrase.
- It does not know about unusual sound labels, so check the result for stray brackets.
- It cannot restore timestamps once removed. Keep the original export if you may need them.
Common routes
If you only need the words, export without time labels using the no-timestamps transcript. If you copied text from YouTube's own transcript panel, paste it into the cleaner. If the next step is a prompt, the long-video guide for ChatGPT explains splitting a cleaned transcript into parts. To keep subtitle timing instead, see transcript file formats.