Clean a transcript in two passes. A mechanical pass removes timestamps, repeated lines and tags, which software does well in seconds. A human pass fixes wrong words, names, punctuation and speaker turns, which software mostly can’t, because it needs to know what was meant. Skipping the first wastes your time. Skipping the second leaves errors that look authoritative because they’re neatly formatted. YouTube’s own help page says to “always review automatic captions and edit any parts that haven’t been properly transcribed.”
What goes wrong in automatic captions
Homophones and near-sounds. Speech recognition picks a plausible word for the sound. In a talk about software, “cache invalidation” can come out as “cash invalidation”. Every word is spelled correctly, so a spell checker won’t help. These are the most dangerous faults.
Names and terms. Product names, people, places and jargon are the words most likely to be wrong, and often the most important to the reader. YouTube notes that captions can misrepresent speech because of “mispronunciations, accents, dialects, or background noise”.
Missing or poor punctuation. Some systems give none, and others guess. Without sentence boundaries, text is hard to read and impossible to split into paragraphs automatically.
Rolling duplicates. Live-style captions repeat each line as it scrolls up. A YouTube auto-caption file, downloaded as VTT, looks something like this:
00:00:00.320 --> 00:00:02.750 align:start position:0%
so<00:00:00.640><c> today</c><00:00:00.960><c> we're</c><00:00:01.200><c> going</c><00:00:01.440><c> to</c>
00:00:02.750 --> 00:00:02.760 align:start position:0%
so today we're going to
00:00:02.760 --> 00:00:05.110 align:start position:0%
so today we're going to
talk<00:00:03.000><c> about</c><00:00:03.300><c> cash</c><00:00:03.600><c> invalidation</c>Stripping the tags alone leaves each phrase two or three times over.
Sound tags. [Music], [Applause] and [Laughter] are useful for viewers and noise in an article.
Speaker confusion. Many systems give no labels. Those that do, such as meeting platforms, often split one person’s turn or merge two people talking over each other. Names come from account display names, so you may get iPhone (2) as a speaker.
Disfluencies. “Um”, false starts and repeated words are faithfully recorded. On the page, speech that sounded fine reads as rambling.
Numbers and formatting. “Twenty twenty six”, “five hundred k”, and “three point one four” can come out as words or digits in no consistent way. Long sentences get cut across cue boundaries with no regard for grammar.
Decide what the transcript is for
This determines how far to go, so settle it first.
| Purpose | Keep | Remove | Effort |
|---|---|---|---|
| Legal, research, or verbatim record | Every word, disfluencies, timestamps | Nothing | Check against audio, don’t tidy |
| Accessibility captions | Words and sound descriptions, accurate timing | Only errors | High, and accuracy is the priority |
| Meeting notes | Decisions, speakers | Most small talk | Low to medium |
| Blog post or article | Meaning | Fillers, tags, timestamps, repetition | Rewrite as prose |
| Searchable archive | Words | Tags and duplicates | Mechanical pass only |
A verbatim transcript shouldn’t go through a filler remover. A research interview may treat a hesitation as data. If in doubt, keep an untouched copy of the original, always.
The workflow
1. Get the best source you can
Download the original caption file instead of copying from a web page. Files keep structure that pages lose. Where the speaker uploaded their own captions, use those instead of the automatic track. Keep this original in a safe place, since you’ll need it to check timing.
2. Run the mechanical pass
Choose the tool for your source:
- YouTube: the YouTube Transcript Cleaner. We gave it a file in the rolling format shown above, and it returned one clean line:
So today we're going to talk about cash invalidation. It removed the header, timing, inline tags and rolling repeats and dropped the[Music]tag. Note what it left: “cash” is still wrong, which is for you to fix. - Zoom or Teams meeting files: the Zoom Transcript Cleaner strips the cue numbers and timing, and keeps speaker names, as in
Priya Shah: Okay, um, let's start with the budget. - Anything else: the general Transcript Cleaner, with options for each step.
Read the first page of output before you go on. Any automated cleaner can do something unexpected with an unusual file.
3. Fix the words, with the audio on
This takes the most time, and there isn’t a shortcut. Play the recording at 1.25 to 1.5 times speed and read along. Listen for wrong words in places that sound right.
Speed it up with a find-and-replace list. Before you begin, write down what you know will be wrong: names, product terms and acronyms. A single Find and Replace pass fixes each one everywhere. Watch out for words that have a normal use too, though. Replacing every “cash” would be a bad idea in a talk on payments.
Spot-check instead of reading everything if the stakes are low. Prioritize numbers, dates, names, quotes and anything someone might act on. An error in a dollar amount is much worse than a clumsy sentence.
4. Sort out speakers
If there are labels, fix the names first. Make them consistent, for example turning iPhone (2) and Tom R into Tom Reed. Then fix turns at the points where people interrupt each other. If you don’t need labels at all, for a single speaker’s talk for example, Remove Speaker Labels strips them. When we tried it, it left a sentence like Here is the thing: it broke alone and removed Speaker 1:. Still, check its results if your text has many sentences with a colon near the start.
5. Add punctuation and paragraphs
If the source has no punctuation, add it by hand while listening, or use a dictation or transcription tool with punctuation restoration to redo the source. Our Paragraph Formatter groups sentences into paragraphs, but it relies on the punctuation that’s already present. It can’t invent sentence breaks. We gave it some text in which the last part had no end marks, and all of that text stayed together as one block.
Break paragraphs at changes of topic, not at a fixed length. A reader wants a new paragraph when the subject moves on.
6. Remove fillers, carefully and last
This goes late because you want to fix meaning before you delete things. A careless filler removal is destructive, because words like “like” and “actually” are often not fillers at all. Consider these:
I like this plan.
It looks like rain.
It was kind of you.Every “like” and “kind of” there has meaning. We ran our Remove Filler Words tool on this:
Um, so I, uh, think we should, you know, ship it. I like this plan. It looks like rain.and got:
So I think we should ship it. I like this plan. It looks like rain.That’s the right behaviour for those sentences, but no rule-based tool gets all cases right, so read the output. Keep a copy of the version before this step. Some fillers carry tone, and a quotation should stay accurate. If you’re quoting someone, don’t change what they said beyond removing obvious stumbles, and follow your publication’s rule on that.
7. Read it as a reader would
Read the whole thing without the audio. Where a sentence doesn’t make sense on the page, go back and listen. Spoken language loops and restarts, and for a publishable article you’ll have to rewrite much of it, not merely clean it. A transcript is raw material. The Transcript to Article tool gives a rough draft structure, with headings taken from the text, and it’s a starting point for editing and nothing more.
8. If captions go back on the video, check the timing
Editing text can throw off timing and line lengths. Netflix’s English timed text guide, a common reference point, sets a limit of 42 characters per line and up to 20 characters per second for adult content. Very long cues won’t be readable in the time they’re on-screen. If you changed a lot, use a subtitle editor such as Subtitle Edit to re-time, as text-only tools can’t do that. For format questions, see SRT versus VTT.
How much time to budget
There’s no universal figure, since it depends on audio quality, the number of speakers and how technical the content is. Time yourself on the first five minutes of the recording, and multiply. Clear single-speaker audio goes quickly. A crosstalk-heavy panel with specialist vocabulary can take several times as long, and at that point it may be cheaper to pay a human transcriptionist than to fix the automatic version.
What software can’t do for you
No cleaner can tell you that “cash” should be “cache”, that the speaker misstated a figure, or that a sarcastic remark will read as sincere on the page. All the tools named here run in your browser on the text you give them. They don’t listen to the audio, and they don’t know what was said. They make the human pass faster. They don’t replace it, and anything published under your name deserves that second pass.