Edit the Words, Edit the Video: Transcript Cleanup 101

Editing video is turning into editing text

Here is a change that is happening quietly. You record yourself talking. A computer turns the audio into a written transcript. Then you edit the video by editing the words. You delete a sentence in the transcript, and the matching piece of video disappears. You take out an "um" in the text, and the "um" is gone from the footage too.

Once you can do that, the transcript becomes your editing timeline. And that means one thing matters more than anything else: how clean the transcript is. A messy transcript makes a messy edit.

So I set out to build a cleanup tool. The goal sounded simple, "get rid of the ums." It turned out to be bigger and more fun than that. This is the whole design, in plain terms, with real examples.

Read a transcript of yourself sometime. You will be surprised. We do not talk in clean sentences. We say "um" and "uh." We repeat words. We start a word, stop, and start over. We say the same thing three times before it comes out right.

Here is a real line from a two-hour recording of mine:

here's the, here's the, uh, here's the, here's the OpenAI API key

Four tries at the same three words. That is normal. Talking is a rough draft you make out loud. The tool's job is to turn that rough draft into something clean to read, without changing what you actually meant.

That last part is the whole challenge. It is easy to build a cleaner that reads great and quietly puts words in your mouth. The hard part is cleaning it up while keeping every real word you said.

The big idea: safe stuff first, smart stuff later

The tool works in stages, cheapest and safest first.

Plain computer code goes first. Not AI, just simple rules. Code is fast, it is free, it runs right on your own computer, and it can never make up a word you did not say. Think of it like spell-check, not like a ghostwriter. Spell-check fixes the obvious stuff and stays out of the way. It does not rewrite your essay.

Then, only for the tricky judgment calls, the tool asks an AI. And after the AI makes a change, a second AI checks the first one's work. More on that later.

What plain code can safely clean up

A few simple rules handle a surprising amount.

Filler sounds. "Um," "uh," "er." These carry no meaning, so they go.

Repeated words. "The the the car" becomes "the car."

Cut-off words. Sometimes you start a word, stop, and say it again: "the th- the main thing." The little "th-" is a false start of "the," so the tool drops it. You can tell it is safe because the stub is the beginning of the word right after it.

Here is a neat part. Watch this sentence get cleaned:

So the th- the main thing is the wor- workshop

First the tool removes the cut-off pieces, "th-" and "wor-." But that leaves "the the" sitting next to each other, which is now a plain repeat. So the tool runs again and collapses that too. The final result:

So the main thing is the workshop

The tool does not clean once and stop. It keeps going in a loop until nothing changes, because fixing one thing often uncovers the next. And since each pass only ever removes text, it always finishes. That loop is how a four-times retake like the OpenAI line above shrinks all the way down to one clean "here's the OpenAI API key."

The mistake that taught me the most

Now the wall I ran into, and it is the most useful lesson here.

I wrote a rule to catch false starts. It looked for a word that repeated and cut out everything in between the two copies. On paper it looked amazing. It removed almost eight percent of the transcript.

Then I actually read what it did. This sentence:

an off-the-cuff teacher, and as much as I try, the best thing I have done

got turned into:

an off-the best thing

The rule saw "the" in "off-the-cuff" and matched it to a later "the," then deleted the whole middle. It ate a real sentence.

The lesson: a smaller cleanup that keeps every real word beats a bigger one that deletes your meaning. I threw the clever rule out. The honest three-and-a-half percent was worth more than the flashy eight percent that lied.

The main rule: two lists, opposite settings

That mistake led to the idea that holds the whole tool together. There are two lists.

One list is for deleting. The other is for flagging, which just means highlighting something for a person to look at.

Here is the trick. These two lists are set up in opposite ways.

The delete list has to be careful, because deleting is permanent. If it gets something wrong, it erases a real word you said. So the delete list only removes things it is completely sure about.

The flag list can be bold, because flagging removes nothing. If it gets something wrong, you just glance at it and move on. So the flag list is allowed to point at anything that looks even a little suspicious.

Same exact situation, opposite response, and the thing that decides is one question: can this be undone? Be careful when it is permanent. Be generous when it is easy to take back. Almost every hard choice in the tool comes back to that.

Need help building something like this? Create a ticket today →

Some corrections wave a little flag

A lot of the time, when you correct yourself, you say a signal word first. "No wait." "I mean." "Sorry." These are like wake words, the phrase you say to wake up a voice assistant. When the tool hears one, it knows a correction is coming and keeps only the fixed version.

It ships Tuesday, no wait, Wednesday

becomes

It ships Wednesday

But some signal words are not trustworthy. The word "not" can mean a correction, but it also shows up in totally normal sentences like "I do not like it." So a strong signal like "no wait" is allowed to delete, while a weak signal like "not" only flags. Strong signal, delete. Weak signal, flag. Same rule as before.

The doubles you cannot judge just by looking

This is my favorite part. "The the" is a stutter, easy. But plenty of doubled words are correct English.

• "Know that that is where it lives." Correct.
• "He had had a long day." Correct.
• "What it is is that it broke." Correct.

The best example came from my real recording:

we're telling it it can crawl

Every instinct says "it it" is a stutter. It is not. It means "we're telling it that it can crawl," with the "that" left out, the way we talk. The first "it" is the thing you are telling. The second "it" is what can crawl. Both are doing a job. Delete one and the sentence falls apart.

No simple rule can tell that apart from a stutter, because the difference is grammar, not spelling. So the tool does not delete these. It keeps both words and flags them for a person to check. When you cannot be sure, you do not guess. You ask.

Two kinds of mistakes

Everything so far is a mistake you can see in the shape of the words. But there is a sneakier kind.

Sometimes the transcription hears a real word, spells it right, and it is still the wrong word. It hears "basic support" as "basis support." Both are real words. The grammar is fine. Nothing looks wrong.

No spelling rule can catch that, because there is nothing misspelled. Only two things can. One is a dictionary of the words that belong in your topic, which knows "basic support" fits and "basis support" does not. The other is an AI that reads for meaning. Mistakes you can see and mistakes you can only understand are two different problems, and they need two different tools.

Let the AI help, then check the AI

When the tool does hand a tricky call to an AI, there is a catch. If you let one AI clean your text, it will quietly rewrite you into itself. It smooths things over and "improves" them until the words are not quite yours anymore.

So the AI never gets the final say on its own work. A second AI, a checker, gets your original words and the AI's version and does one job: point out anything that was added, removed, or changed in meaning. It is a reviewer reviewing the reviewer.

And that checker is not just for show. I once wrote a rule to fix swapped words, like when you say "in, at my job" and mean "at my job." It ran on the real recording, fired sixteen times, and was wrong twelve of those times. It turned "talk to, of course" into a broken phrase. That rule got demoted from deleting to flagging the same day. An unsure fix is a guess, and guesses get checked.

Three ways to see it

The tool shows every transcript three ways. One view paints each kind of change its own color so you can study it. Another view looks like a code diff, red for what got removed and green for what got added. And the last view is the clean read-through, with the removals gone and the fixes applied, so you can read it start to finish and hear how it sounds.

What it turned into

I started out trying to delete some "ums." I ended up with something that has to know that "um" is noise, that "that that" might be perfectly good English, that "not" might mean scratch-that, and that the "like" in "smells like flowers" is a completely different word from the "like" in "like, whatever."

That is not really about deleting words anymore. It is about teaching a computer to listen the way a careful editor listens. And once the transcript is clean, most of the video edit is already done. The words are the timeline now.