PricingEveryone says “um.” Most editors make you find each one, cut it, and then fix everything that was timed to the speech. Shorty now does it inside the edit. Write “remove the ums” in your brief, and the director cuts the hesitations, stutters and false starts, keeps every word that carries meaning, and builds the rest of the edit on the clean take, captions included. Below are four real talks, before and after, with every cut marked.
It’s One Sentence in Your Brief
There’s no separate tool to open and nothing to export first. Add a line to your brief, like “remove the ums and false starts,” or ask for it in a revision of a finished cut. The director listens to the speech, decides what is a filler in context, cuts it from the picture and the sound, and keeps editing.
That matters because removing fillers changes the timing of everything after them. Cut 28 hesitations from a two-minute answer, and every caption, sound cue and B-roll shot after the first cut is late. In a standalone filler tool, you fix that by hand or start your captions over after exporting. In Shorty, captions, sound cues and the edit’s length are rebuilt from the cut transcript, so nothing drifts.
Hear It on Four Real Talks
Pick a talk and press play on either side. Click any word to hear it, or use Hear a cut to play the same moment before and after. These are public talks on YouTube, linked under each one, run through the same filler-removal tool Shorty’s director uses inside an edit. The speakers aren’t affiliated with Shorty.
Nik Storonsky on building Revolut
Nikolay Storonsky, Revolut, at the Singapore FinTech Festival · Source on YouTube, from 8:10
- Length
- 2:00 → 1:49.8
- Words cut
- 40
- Cuts
- 28
- Ums and uhs left
- 0 of 13
Uh, third thing is, uh, basically that we have B2C clients as well, right? What it means in practice, uh, we, we effectively can help your, uh, to acquire funds from our own customers for free. And because it's within our system, so if, for example, I'm a Revolut user, I'm going to check out on Revolut mission, transaction- Got it is instant and free. So that's another thing. Fourth thing that we are going to offer you is effectively to, to, uh, sell direct to our retail customers, to become a sales channel as well. So we have all these natural advantages that B2B businesses like Stripe and Adyen don't have. As a result- Got it I think we can compete with them and we can of- offer much better price product and much co- more competitive as well. Yeah. That's, that's, that's fascinating. B- so mo- more broadly, I mean, you know, you, you're launching products that, uh, uh, as I said, entire companies are built on, and, and you've been doing that continuously, um, and at pace. Those are, you know, high quality products. So I'd be curious to hear, and I'm sure a lot of people, you know, entrepreneurs, uh, on, on, on this, uh, listening to us would love to hear what is your secret sauce and how you organize yourself, especially on the product and, and, and tech side, to be able to keep on running so fast with such a wide range of, of, uh, of products? So basically, if you look at every single company with a single line product, uh, so the way it works, maybe 20, 30% of the company actually works on the product itself, right? While the, the rest 70% is supporting this product. W- what I mean by supporting is your f- financial services, right? It's your fin crime services, your compliance services, your legal services, right? HR people and so on. Then the way we build, uh, the whole org or the system within the org, is that you have all these services supporting many businesses, right?
The busiest talker of the four. Notice “B2B businesses like Stripe and Adyen”: that “like” means “such as”, so it stays.
Prime Intellect on inference at scale
The Prime Intellect team at an NVIDIA developer event · Source on YouTube, from 20:00
- Length
- 1:00 → 0:53.2
- Words cut
- 38
- Cuts
- 27
- Ums and uhs left
- 0 of 16
Some of you may be TensorRT. Uh, but like when you try to scale your inference, you realize that like the bottleneck isn't really like the engine and the, the kernel implementation there. The bottleneck becomes like the whole orchestration layer around that, of like if you're multi- uh, if you're doing like a data parallel thing where you have like multiple, uh, engines running, then, uh, then you have like first issue of like being, uh, like cache aware. Of if I'm doing like a multi-turn thing where my, my agent is doing multiple tool calls and then coming back into the inference, uh, I want to make sure that it hits the same engine and it's able to hit that prefix cache. And also, uh, PD disaggregation concerns, where you might have a prefill node that runs a config that is better for throughput, but then you have like a, a, a decode node that runs better for latency, and then how do I, uh, sh- uh, do the KB cache transfer between that? Uh, and all the orchestration, auto-scaling kind of stuff. And so that's what we use N- uh, NVIDIA Dynamo for. Uh, we also have, uh, GPU and CPU sandboxes. Uh, I think GPU sandboxes right now is in beta. Uh, our research team is trying them out, but I think we're gonna release them in a few weeks. Uh, for very-
Here every “like” was a verbal tic, so all 12 went, along with 16 ums and uhs, a stutter and two abandoned words.
Arjun Karanam on continual learning
Arjun Karanam, Trajectory, at a Sequoia Capital event · Source on YouTube, from 10:57
- Length
- 2:00 → 1:54.4
- Words cut
- 22
- Cuts
- 21
- Ums and uhs left
- 0 of 8
Locks the door to owning, owning your weights and then continually improving on top of them. Um, and then Gabe also talked about this, this idea that, like, model routers are, uh, going to probably play a really large role in routing intelligence to the exact capability of the task they're at. And so experiment with, with routers is another wish I would wish cast into the world. And so that is the high level here. And so I think, like, when we work with companies, the first thing we often do is do an audit of, of these and, like, where things are. And for the most part, there's a lot of work to be done to get here, and so we work with them to build it out. And it would be my dream if we go to a company and they have a lot of this built out. Granted, a lot of the work we're doing is also knowing that a lot of this we can't expect, and so we're doing a lot of cool research on, you know, assuming that nobody has evals, somebody doesn't have evals, how can we overcome that? Or assuming that the traceability is off, like, how can we, uh, learn from the signals that are there? And so that is, that, that is my, my goals for, for the universe here. Uh, just wanna quickly go over, uh, I think this is, like, a very, very core research and product problem. And so we're really excited with the team we've put together. But I think more importantly, and the two of the folks who are here today, uh, we're working with extremely frontier customers, the folks who are really, really pushing the boundaries of what agents can do in production both today and tomorrow. Uh, and so really excited to be working with them. And another thing I wanna show is we have... Oh, that's my email. Uh, we have a beta. And so a really, really core belief we have is that this capability, this ability to own your own intelligence should not be something that you need to consult away a lot of the times. It should be expertise you build on your own because-
A conference talk over slides: eight ums and uhs, five tic “like”s and a few restarted phrases.
Elon Musk at the Tesla Roadster event
Elon Musk on stage at Tesla’s Roadster event · Source on YouTube, from 16:00
- Length
- 2:00 → 1:58.4
- Words cut
- 7
- Cuts
- 4
- Ums and uhs left
- 0 of 1
Go to zero to 60? I can't remember. Well, you gotta break one second. Obviously. I tested this the other day. Can we put it up, guys? Can we put the test up? Like, this, this goes fast. We wanted... All right. Woo! One second. Right? You said one second. Well, w- I said less than one second. Oh, shit. Hey, guys. We have a car out there, right? Don't we have a car out there? Yeah! Can you put it on the max? Just see- let's try and beat one second. Do we have a time? What was the time, guys? We didn't have a whole lot of room out on the pad, so we had to cut it short, but do we have a live time? I'm waiting for it. Well, while we're getting the live time, uh, you know, I will say this. Elon mentioned the Rod-
A mostly clean stretch on a loud stage. Only seven words came out, because there was little to cut.
Struck through words were removed. Everything else is the speaker’s own audio, untouched. Across the four talks, 107 words came out in 80 cuts, including all 38 ums and uhs in the transcripts, and 7:00 of speech became 6:35.8.
What Counts as a Filler, and What Doesn’t
Ums and uhs are the easy part. The hard part is everything that sounds like a filler but isn’t always one. “Like” can be a tic or a comparison. “You know” can be a habit or a real question. A repeated word can be a stutter or emphasis. A keyword list deletes every match. Shorty’s director reads each candidate in its sentence before it cuts.
In the Revolut clip, it removed “basically,” “I mean,” three “you know”s and every “uh,” but kept “B2B businesses like Stripe and Adyen,” where “like” means “such as.” In the NVIDIA talk, all 12 “like”s were tics, and all 12 went.
Stutters and false starts work the same way. “That’s, that’s, that’s fascinating” keeps the last “that’s.” “If you’re multi- uh, if you’re doing a data parallel thing” keeps the finished sentence.
What It Cuts
Hesitations
Um, uh, er and the long “uhhh” between thoughts.
Stutters and repeats
“The, the,” “my, my,” “to, to.” The last, complete one stays.
False starts
A sentence abandoned halfway and started again. The finished version stays.
Verbal tics
“Like,” “basically,” “you know,” “I mean,” only where they carry no meaning.
A slow start
Tell it where the take really begins, and everything before that line goes.
How It Works Under the Hood
It starts with a transcript that doesn’t tidy up. Most speech recognizers are trained to produce clean, readable text, so they quietly drop the ums and stutters, which are exactly the words you want to cut. When we ran a widely used transcription model on six interview excerpts with audible hesitations, its transcripts contained zero “uh”s and zero “um”s.
So Shorty now uses a verbatim recognizer built for more than 90 languages. It writes down every hesitation, false start and repeated word, gives each word its own timestamp, and labels laughter and applause as sounds rather than words, so a laugh is never mistaken for speech. The transcript is made once per clip and shared, so your captions and your cuts are planned on the same words.
Recognizer timestamps are good enough for captions but not for surgery. So Shorty also runs a phonetic forced aligner on the original audio, which pins each word, and each sound inside it, to the waveform. Cut points come from those sound boundaries. If a cut would sit next to a word the aligner can’t place with confidence, it is refused rather than risked.
The cut happens in the original recording at 48 kHz, with a 4-millisecond equal-power crossfade at each join, so there’s no click and no gap. Picture cuts land on whole frames, and the audio is muxed with its encoder delay accounted for, so lips stay in sync. Shorty never cuts through a word it didn’t choose, never removes more than 30% of a recording as filler, and checks before delivery that every kept word survives in order.
On the four talks above, every kept word survived in order, the audio between the joins matched the source sample for sample, give or take one step of rounding, and sync measured zero samples of drift at 5, 20 and 40 seconds into each cleaned clip.
How It Compares
Most editors now have a filler-word button. Here is how the workflows differ, based on each tool’s own documentation.
| Tool | How you start it | What it looks for | What happens next |
|---|---|---|---|
| Shorty | One sentence in your brief or in a revision | Ums and uhs, stutters, repeats, false starts, and “like”, “basically” or “you know” only where they carry no meaning | The whole edit, with captions, music cues and length, is built on the clean take |
| Descript | AI Tools → Remove filler words, then review the list | “Um”, “uh”, “you know”; retakes and long gaps are separate tools | Captions follow the edited script; an “Avoid harsh cuts” option skips fillers that would clip nearby words |
| CapCut | Auto captions with “Identify filler words” on, then delete the highlights | “Um”, “uh”, “like”, “you know” and pauses, which CapCut calls examples | Transcript edits are linked to captions; you finish the edit yourself |
| Premiere Pro | Transcript panel → filter by Filler words → Delete all | “Uh” and “umm” | A timeline you keep editing by hand |
| Riverside | AI tools → Remove filler words → review each card | “Um”, “uh”, “hmm”; “like”, “you know”, “I mean” when used as fillers; breaths | A separate “Harsh cuts” group flags cuts that can clip nearby words |
| VEED | Magic Cut, one click | Silences, ums and ahs, repeated mistakes, awkward pauses | English only, according to its help centre |
| OpusClip | Speech Cleanup in the editor | “Um”, “uh”, “you know”, stutter repeats; bad takes on Pro | Captions added before cleanup shift and need redoing, according to its API guide |
Sources: Descript · Descript help · CapCut · Premiere Pro · Riverside · VEED · OpusClip · OpusClip API guide. Checked 20 October 2026. Plans and features change, so check each tool’s current page.
Every tool here can find an “um.” The difference is what happens around it. A filler button is a pass you run, review and export before the real editing starts, and anything timed to the speech has to catch up afterward. When we asked ChatGPT how to clean up a talking-head video for Reels, it suggested removing the fillers in one app, exporting, and adding captions in another. In Shorty, that’s one brief.
Ask for It Your Way
“Make a 45-second vertical cut of this interview. Remove the ums, uhs and false starts, keep my pauses natural, and add word-by-word captions.”
“Only cut the ums and uhs. Leave “like” and “you know” in; that’s how I talk.”
“Start the video on “Here’s the thing” and cut everything I said before it.”
“Remove the filler words from the second clip and keep everything else the same.”
Go Deeper
Filler-word removal in Shorty
What you provide, what Shorty decides and what you get back.
How to remove filler words from a video
The one-prompt way, and the steps in Descript, CapCut and Premiere Pro.
Remove filler words without choppy cuts
Where to cut, how to crossfade and how to hide a jump.
Remove filler words and add captions in one step
Why the order matters for caption timing.
Best AI tools to remove filler words
Ten tools compared from their own documentation.
What are filler words?
Which ones to cut, which to keep, and how often speakers use them.
Frequently Asked Questions
How do I remove filler words like um and uh from a video automatically?
In Shorty, add your footage and write it in your brief, for example “remove the ums, uhs and false starts.” The director finds the hesitations, stutters and restarts in the speech, cuts them from the picture and the sound, and builds the rest of the edit, captions included, on the clean take. You can also ask for it in a revision of an edit you already made.
Does it remove every um?
It removes every hesitation its transcript records. On the four talks in this post, that was all 38. A hesitation the recognizer doesn’t hear can’t be cut, so listen to the result, and ask in a revision if one slipped through.
Will it cut a “like” I meant to say?
It shouldn’t. Words such as “like,” “basically” and “you know” are judged in their sentence and only cut when they carry no meaning. In the Revolut clip, “businesses like Stripe and Adyen” stayed. If you want them all kept, say so in your brief.
Will the cuts sound choppy?
Each join is cut at a sound boundary in the original audio with a 4-millisecond crossfade, and pauses that aren’t part of a cut are kept. On camera, a cut can still be visible as a small jump. Ask for B-roll or captions over those moments in the same brief.
Do my captions still line up after the cuts?
Yes. Captions are rebuilt from the cut transcript, so every word lands where it is now spoken. The same goes for sound cues and the edit’s length.
Which languages does it support?
English and Russian are tested by ear today. French, German, Spanish, Portuguese, Italian and Dutch are set up and in testing. Speech that mixes languages isn’t supported yet.
How long can a recording be?
Up to three minutes of speech per clip, and up to four clips per edit, which suits short-form talking heads, interviews and Reels.
How is this different from Descript or CapCut?
Those tools make filler removal a separate pass. You run it, review it, export, and then edit. In Shorty it is part of the edit: one sentence in your brief, judged in context, with captions, music and length rebuilt on the clean take.
Talk the way you talk. We’ll take out the ums.
The Shorty Team
Say it once. Cut the ums in the same edit.
Add your talking-head footage and write “remove the ums” in your brief. Captions, music and length follow the clean take.
Start creating ↗