Remove filler words, pauses and false starts without flattening your voice

Cutting ums out of a talking video is the least creative job in editing and the one that eats the most time. It is also the job where automation is closest to solved and still gets it wrong in specific, predictable ways. This is what each tool actually detects, the six places automatic cutting damages speech, and the settings that keep breaths where they belong.

Four things people call one thing

The first and third rows are what a one-click cleanup is genuinely good at. The second and fourth are where the review pass earns its keep, and the last two are where a badly configured tool does real damage that nobody can name afterwards. The video just feels tiring.

Manual against automatic, honestly

At around 150 words a minute, the research rate puts roughly nine or ten decisions in every spoken minute, before you count silences. That is the number that makes a daily posting habit collapse: not the filming, not the captions, but nine small decisions a minute, every day.

What each tool actually detects

Two details worth knowing before you buy anything. Descript's help centre says filler word detection supports English transcripts only, and that its transcription is limited to Latin-alphabet languages. Adobe's position is the opposite: filler detection rides on the acoustic pattern rather than the word, so it claims to work across every language Speech to Text handles. Both are September 2026 statements from the vendors themselves.

Where automatic cutting breaks speech

1. It eats the inhale. Set the silence threshold at 150-200 ms and every breath is longer than the threshold, so it goes. Sentences start butting into each other and a viewer gets tired forty seconds in without knowing why.
2. It cuts on a word boundary instead of a zero crossing, and you get a click at the join. Plosives at the start of the next word make it worse.
3. It removes a word that was doing work. 'So, here is the thing' and 'so it doubles' are not the same 'so'. A word list cannot tell them apart.
4. It leaves half a false start. You said 'I want... I am going to show you', the tool removed the gap and glued the two halves into nonsense.
5. It flattens a list. The pauses between items in a list are the punctuation of that list. Remove them and the whole thing becomes one breathless run.
6. It kills a rhetorical beat. The pause after your own question, before you answer it, is a device. A threshold has no idea.
7. It desynchronises captions, if you cut after generating them. The order is always: clean the audio, cut the speech, then caption.

Settings that keep breaths natural

• Minimum silence length: 400 ms or more. Breath research discards pauses under 300 ms when isolating breath sounds because a breath is usually longer than that, so any threshold below 300 ms is cutting into breathing.
• Padding around each cut: 80-120 ms on each side. That is an observation rather than a standard, but under about 100 ms you start hearing clicks on consonants.
• Maximum remaining pause: 250-300 ms. Shorten long pauses to this rather than closing them completely, or the delivery turns into a stream.
• Never touch the breath before an important line. In speech-synthesis work, a model trained with detected breath marks scored 3.55 for naturalness against 3.34 without, with real recorded speech at 3.92. Breathing is part of what makes a voice sound human.
• Level threshold: -40 dB or lower. At -30 dB you start losing the quiet ends of sentences.
• Denoise before you detect silence. On a noisy recording the silence detector finds no silence at all, which is covered in our guide to removing background noise.

A ten-minute pass on one take

1. Denoise first, then get a transcript with word-level timestamps.
2. Cut false starts before anything else. They are the longest items and the biggest win on runtime.
3. Run the automatic filler pass, then review every proposed cut rather than accepting the batch.
4. Shorten the remaining pauses to 250-300 ms instead of closing them.
5. Listen back at 1x with the picture off and fix the joins your ear catches.
6. Only now generate captions, so the timings match the final cut.
7. Keep a note of your own filler words. Everyone has three or four, they repeat in every video, and knowing yours halves the review time on the next one.

The cheaper fix is upstream: fewer stumbles going in. Reading from a prompter removes most false starts because you are not searching for the next thought while the camera runs, which we cover in how to read a teleprompter. For the other passes of a talking-head edit, see how to edit talking-head videos.

How Monty handles it

Monty does this pass as part of building the video rather than as a separate chore. You upload an MP4 or MOV take from your phone, or write a script and record it with the free browser teleprompter. Monty cuts pauses and false starts, times captions to the speech and keeps them inside the safe area, places music under the voice, picks b-roll by meaning, levels loudness, applies your branding and reframes to vertical around your face. You get an editable preview before the render, so a join you disagree with is still yours to change.

From there it publishes to YouTube Shorts, Instagram Reels, TikTok and Telegram with an approval step and a per-platform title and description; Monty is in closed beta, so publishing access is opened per account while the beta runs. What it does not do: generate AI avatars or text-to-video scenes, or build faceless stock videos from a prompt. The input is always a take you recorded.