I counted the "um"s in a 52 minute interview last week. There were 411 of them. Plus 96 "like"s used as padding, 38 "you know"s, and one host who said "right?" at the end of 60 separate sentences. If I had cut every single one, the episode would have lost about four and a half minutes. It also would have sounded like a hostage video.
That is the trap. Filler removal looks like free time savings on a spreadsheet and it is not. Every filler word sits inside a breath, and breath is what makes a person sound like a person. Cut the word, keep the breath, and you get a strange little pocket of air. Cut the word and the breath together, and you get a jump in the room tone that the ear catches even when the eye does not.
My position, after a few hundred episodes: filler words are not the problem. Dead weight is the problem, and filler is only sometimes dead weight.
What actually makes audio sound robotic
People blame the cuts. It is almost never the cuts on their own. It is three things stacking up.
The first is rhythm. Human speech has a natural beat that shifts with emotion. When you strip out every hesitation, the beat becomes metronomic. Sentences start landing at even intervals, and the brain reads even intervals as machine generated. A listener cannot tell you why the episode feels off. They just close the app around minute nine.
The second is room tone. Your voice has silence attached to it, and that silence is not empty. It has the hum of the AC, the shape of the room, the noise floor of the mic. Chop a word out of the middle and you often chop out a slightly different flavour of silence than what sits on either side. Stack forty of those and the background starts to flutter.
The third is breath. New editors delete breaths because breaths look like noise on a waveform. Breaths are not noise. A breath before a sentence tells the listener something is coming. A breath after a hard statement tells them it mattered. Remove them all and the host sounds like they have no lungs.
A filler word is a symptom. Silence detection treats it as a defect. A human editor asks what the speaker was doing in that half second, and most of the time they were thinking, which is the whole point of an interview.
Why we do not run silence detection
This is the part where I get slightly unpopular with people who love a one click plugin. We do not auto cut by silence detection. Every episode we take on gets watched end to end by a human editor. Not scrubbed. Watched.
The reason is simple. An algorithm sees a 900 millisecond gap and calls it dead air. It cannot tell that the gap came after a guest was asked about their father, and that the pause is the single best moment in the episode. It cannot tell that a stammer is a host getting genuinely excited rather than losing the thread. It cannot tell that "I mean" is filler in one sentence and a real correction in the next.
Machines are excellent at finding candidates. They are terrible at deciding. So we let a person decide, every time, and we charge accordingly. Starter runs $199 per episode for a single camera 1080p show with one revision. Studio runs $299 for two to four cameras in 4K with unlimited revisions. Both cover up to 60 minutes as standard, 30 minute shows come in cheaper, 90 minute shows cost more. If you are running weekly, a 4-episode pack saves 15% and going monthly saves 25%. I would rather price the human in than sell you a faster robot.
The three tier system I use on every pass
Here is the actual working method. I sort every filler instance into one of three buckets before touching anything.
Tier one, cut without thinking. These are the ones that carry zero information and sit at a clean boundary.
- Stacked repeats at the start of a sentence: "So so so, the thing is"
- False starts that get immediately corrected: "We went to, we flew to Berlin"
- Throat clearing and lip smacks before a clean word
- "Um" sitting alone between two complete sentences with a real pause on both sides
Those are boundary cuts. The word is bracketed by silence you can trim symmetrically, and nothing about the rhythm changes.
Tier two, cut only if you can hide the seam. Filler that sits mid sentence, inside the flow of a phrase, where the words on either side are running into each other. "The thing about, um, growth is." You can take that out, but you have to leave a sliver of the original gap or the two halves collide.
My rule is to leave 60 to 120 milliseconds of the original room tone in place. Not silence. The actual air from that exact moment. It sounds like nothing and it does everything.
Tier three, leave it alone. Filler that is doing emotional work.
- A hesitation before a hard admission
- "You know" used to check in with the guest, which is a social move, not a verbal tic
- Any stammer during a genuinely emotional beat
- Filler in the first 30 seconds of a cold open, because a slightly rough opening reads as real and a polished one reads as an ad
That last one surprises people. I have gone back and forth on it and I keep landing in the same place, which is why I wrote up my thinking on how long a cold open should run as its own piece. Perfectly clean opens test worse than slightly human ones. Every time.
The room tone rule, in practice
If you take one technical thing from this post, take this. Never cut to hard digital silence. Ever.
When you remove a word, you have created a hole. If you close the hole completely, the waveform slams shut and the noise floor drops to nothing for a few frames. The human ear is remarkably good at hearing the absence of a noise floor. It registers as a click or a swallow even when no actual click exists.
The fix is boring and it works. Grab a clean two or three second stretch of the host breathing and not talking, usually right at the top before the episode starts. That is your room tone bed. Any time a cut leaves a gap that needs filling, drop a fragment of that bed underneath. The seam vanishes.
On a two to four camera setup this matters even more, because you are usually cutting picture at the same time and the eye will forgive a lot that the ear will not. Our multi-cam editing passes are built around audio first for exactly that reason. Get the audio timeline right, then hang the picture on it.
How many should you actually cut
I get asked for a percentage and I refuse to give one, because the honest answer depends entirely on the show.
A narrative show with a scripted host, cut hard. That voice is meant to sound composed, so a 90% removal rate is fine and the polish is the product.
A two person conversation, cut maybe half. You are preserving the feeling that two people are actually talking rather than performing.
A solo founder or expert talking through something technical, cut less than you think. Hesitation reads as thinking, and thinking reads as credibility. A single camera show where somebody explains their own work is more persuasive with the seams showing. It runs the other way for music shows, where hosts are used to hearing takes tightened, so our podcast editing for musicians work tends to sit a notch cleaner than the average interview.
An emotional or long form interview, barely cut at all. Trim the stacked repeats and stop.
The related question is when to remove entire passages rather than words, and that is a different skill with different rules. I put down the framework in when to cut a tangent and when to keep it, because deciding what stays is genuinely harder than deciding how to cut it.
Where the robot sound sneaks back in
Two places, and both are late in the process.
Clips. The vertical clip format punishes any awkwardness because there is nowhere to hide. So the temptation is to cut a 45 second clip down to 38 seconds of pure statements, and what you end up with is a person barking. Our vertical clips run $25 each as an add-on and I always tell people to let the clip breathe by one beat more than feels right. The clip lives or dies on whether the speaker sounds human in the first two seconds. Same logic applies to how we handle clips as part of a full episode delivery, and if the clip is heading to YouTube, packaging carries just as much weight as the cut, which is why our sister studio ClickTheory exists at all.
The second place is show notes and chapters. When you write chapter titles from a transcript that has had all the filler stripped, the titles come out sounding like corporate headings. Read the raw transcript instead. Show notes and chapters run $35 with us and the good ones use the host's own phrasing, warts included.
What this costs you in time, and why that is fine
A careful filler pass on a 60 minute episode adds somewhere between 40 and 90 minutes of editor time compared to running a plugin. That is the real number. It is the difference between an episode that sounds edited and an episode that sounds like it was never touched.
We still turn the first cut around in 48 hours, and revisions come back the next working day. If something needs to be out tomorrow, rush 24 hour delivery is $75. The time cost lands on us, not on your publish schedule, which is how it should be.
If you are weighing whether to do this in house or hand it over, the honest maths is in how agencies price podcast work. And if you are handing files to anyone at all, sort out contracts and file ownership before you send the first drive, not after.
The test I run before I sign anything off
I play the episode at normal speed, eyes closed, from a cold start. Not scrubbing. Not at 1.5x. Just listening like a listener.
If I flinch at any point, something is wrong at that timestamp. Nine times out of ten it is a cut that is 40 milliseconds too tight, or a breath that should not have gone. I fix it and run the pass again from the top.
That is the whole quality gate. It is not clever and it does not scale to a plugin, which is exactly the point. If you want to hear what this sounds like on your own show, send me an episode at samy@pixelsamy.com and I will cut you a sample. I am in Dubai and work with hosts everywhere, so the timezone is not an issue.
Cut the filler that is hiding your point. Keep the filler that is proving you are a person. That is the entire job.