How to Remove “Um” and “Uh” From Audio (Without Cutting Real Words)

To remove “um” and “uh” from audio, you need to find each filler word, cut it with a little padding on both sides, and crossfade the join so the edit isn't audible. Doing that by hand in a waveform editor works, but it's slow. Doing it automatically works, as long as you can review each cut before it's applied.

What counts as a filler word?

A filler word is a sound or short phrase that occupies a pause without adding information. Linguists call the pure-sound kind disfluencies: “um,” “uh,” “er,” “hmm.” A second family are discourse markers, ordinary words that sometimes act as fillers: “like,” “you know,” “I mean,” “so,” “actually,” “basically.” The difference matters for editing: the first family is almost always safe to delete, the second is only a filler depending on the sentence.

Fillers aren't a defect in themselves. Everybody uses them, and a small number keep speech sounding conversational. The problem is density. When a listener starts noticing them, they're competing with your point. Our filler words list sorts the common ones into always-cut, sometimes-cut and keep.

Why silence removal alone doesn't fix filler words

A silence remover looks for quiet. An “um” isn't quiet. It's a voiced sound with a steady pitch and volume, which is exactly what a voice-activity detector classifies as speech. So a pause cutter will tighten the gaps around it and leave the filler word untouched. Fillers need their own detection step, usually based on a transcript or a model trained to recognise them.

The two jobs are complementary. Trimming dead air tightens the rhythm; removing fillers tightens the language. Most creators want both, which is why the order matters: remove fillers first, then collapse the pauses they leave behind. If you do it the other way round, you can end up with an awkward half-second gap where an “um” used to be. See how to cut silence without a choppy sound for the padding and crossfade settings.

Three ways to remove filler words, compared

1. By hand in a waveform editor

Listen through, select each “um,” delete it, and smooth the join with a short crossfade. You get full control and the cleanest result when you're careful. The cost is time: expect it to take as long as the recording itself, often longer. Fine for a five-minute clip, painful for an hour of podcast.

2. Transcript-based editing

Tools that transcribe your audio let you delete words from the text and cut the audio to match. It's fast, but accuracy depends on the transcript. Some transcription engines quietly drop fillers before you ever see them, which means they can't be selected and cut. Check that your tool keeps them in the transcript, otherwise the “um”s you can hear will be invisible in the text.

3. Automatic detection with review

The model finds every candidate filler and shows you the list. You approve or skip each one, then the app applies the cuts with padding and crossfades. This is the approach NoiseZero takes on Android, and it's the best balance of speed and safety: the machine does the searching, you keep the final say.

Step by step: remove fillers without damaging the audio

1. Start with a clean source. Noise and room echo make detection less reliable. If the recording is noisy, reduce background noise first.
2. Run detection and read the list. You want to see each filler grouped by word with a count, so a quick scan tells you what the tool believes it found.
3. Approve the safe ones. Tick the pure hesitation sounds (um, uh, er). Leave ambiguous words off until you've listened.
4. Check the joins. Each cut should be padded and crossfaded. If a join sounds rushed, restore a short pause rather than cutting it flush.
5. Collapse leftover dead air. After fillers are gone, shorten the long gaps, but keep a quarter-second or so so the speech still breathes.
6. Listen to the whole thing once. Edits are easiest to catch on a straight run-through at normal speed, not when scrubbing.

What to check before you accept a cut

• Is it really a filler? “Like” and “so” are fillers in some sentences and real words in others. “I like it” must survive.
• Does the sentence still flow? Removing an “um” between two clauses can leave a rushed join. A short pause left in place often sounds more natural than a hard cut.
• Is the cut padded? Cutting exactly at the edge of the sound clips the neighbouring consonants.
• Did the speaker stress it? A word carrying emphasis usually isn't a filler, even if it's on the usual list.

Common mistakes

• Deleting every pause along with the fillers. This produces the rushed, robotic sound that people describe as “choppy.”
• Bulk-accepting ambiguous words. Removing “like” everywhere will eventually cut a real one.
• Editing without listening back. A cut that looks clean on the waveform can still click.
• Ignoring other languages. If you mix languages, English-only detection will miss half of your fillers. See the Hinglish filler words list for an example.

How much is too much?

There's no magic number, but a useful test is whether you notice them when you listen back at normal speed. If you do, tighten. If you can't, leave it alone. To measure your own habit, paste a transcript into the filler word counter. It runs in your browser with nothing uploaded.

Do it from your phone

In NoiseZero, record or import a file (M4A, MP3, WAV, AAC, OGG or Opus, up to 60 minutes), let it find the “um”s and “uh”s, approve the cuts you want, and share the result straight to WhatsApp, Telegram, Drive or Gmail. Silence cutting and on-device noise reduction are free; AI filler removal uses optional AI Minutes packs. Audio sent for filler detection is deleted right after processing.

Frequently asked questions

Can AI remove “um” from audio automatically?

Yes. A detection model flags each filler, and the app cuts it with padding and a crossfade. The reliable versions let you review the list before anything is removed.

Will removing fillers make my voice sound unnatural?

Only if every pause goes with them. Keep short natural pauses and crossfade each join, and the result sounds like a tighter take rather than an edited one.

Does removing fillers work for languages other than English?

It depends on the tool. English-only models miss fillers in other languages. NoiseZero currently supports English for filler detection, with Hinglish, Tamil and Telugu built and next on the roadmap.

Is my audio uploaded?

Silence cutting and noise reduction on device stay on your phone. Only the optional AI filler detection sends audio, and it is deleted immediately after processing.