A few filler words per minute is normal and does not hurt how you come across. In a 2024 study, listeners rated speeches with 5 fillers per minute no worse on effectiveness than clean ones, while 12 per minute lowered their ratings, especially when the fillers were sounds like um and uh. This post covers what that research measured, why people say um at all, how to count your own rate, and how to decide which fillers to cut from a recording.
What the 2024 study tested
Matthew Laske and Florence DiGennaro Reed published “Um, so, like, do speech disfluencies matter? A parametric evaluation of filler sounds and words” in the Journal of Applied Behavior Analysis in 2024 (volume 57, issue 3, pages 574 to 583, doi:10.1002/jaba.1093).
They created speeches with filler sounds and filler words at four rates: 0, 2, 5 and 12 per minute. Listeners recruited through a crowdsourcing service for survey research then rated each speaker’s public speaking performance. The abstract reports these results:
| Rate per minute | What the abstract reports |
|---|---|
| 0 | Baseline, no fillers |
| 2 | Tested; no separate finding called out in the abstract |
| 5 | Did not adversely affect perceived effectiveness |
| 12 | Perceptions affected, notably for filler sounds |
The authors conclude that reducing filler sounds matters for perceived effectiveness, and that “five or fewer disfluencies per minute may be acceptable.” Note the wording. It is a cautious conclusion from one study with created speeches and online raters, not a fixed rule. A conversational podcast or a lecture may be judged differently from a prepared speech.
Filler sounds and filler words are different
The study separated two kinds of filler:
- Filler sounds, such as um, uh and er. They are not words in the usual sense.
- Filler words, such as like, so and you know. They are real words used without adding meaning.
According to the abstract, the effect on perceptions was strongest for filler sounds. Tom Critchfield’s summary on the Association for Behavior Analysis International blog reads the same pattern across the data: non-word disfluencies hurt listener appraisals more than word disfluencies.
For editing, that gives you an order of work. Cut um and uh first. Treat like and so with more care, because they are harder to remove cleanly and sometimes carry meaning.
Why people say um and uh
Herbert Clark and Jean Fox Tree studied this in “Using uh and um in spontaneous speaking,” published in Cognition in 2002 (volume 84, pages 73 to 111). Working from several large collections of recorded spontaneous speech, they proposed that speakers use uh and um to announce a delay they expect: uh for a minor delay, um for a major one.
They also argue that listeners can read more into these signals. A speaker may be searching for a word, deciding what to say next, wanting to keep the floor or wanting to hand it over. Their conclusion is that uh and um are conventional English words, planned and produced like any other word.
That is useful context for editing. A filler is often doing a job. In an interview, an um can tell the other person you are still holding the floor while you think. Remove it and you may leave a gap that sounds stranger than the um did.
How to count your own rate
You do not need software to measure this. You need a recording of you speaking naturally and a few minutes of attention.
- Record 3 to 5 minutes of yourself explaining something you know well, without a script. A real call or podcast segment works better than a test recording.
- Note the exact speaking time. Leave out long stretches where someone else talks.
- Listen back, or read an auto-generated transcript, and tally filler sounds and filler words in separate columns.
- Divide each tally by the number of minutes.
| Example | Filler sounds | Filler words | Minutes | Rate per minute |
|---|---|---|---|---|
| Before practice | 22 | 9 | 4 | 5.5 sounds, 2.25 words |
| After practice | 10 | 7 | 4 | 2.5 sounds, 1.75 words |
The numbers in this table are an illustration of the arithmetic, not data from a study.
A transcript is faster than listening, but check that it actually writes the fillers down. Premiere Pro, for example, only labels fillers in transcripts made after the feature was added, so an older transcript would make your count look better than it is.
When to keep some fillers
Zero fillers is not the goal, and an over-cleaned recording can sound edited.
- Conversation. In a two-person podcast or interview, fillers help mark who is holding the turn while they think. Keep a few.
- Meaning-bearing words. “Like” in “it looked like rain” is not a filler. Auto-detection can mark it anyway, so review every cut.
- Rhythm. Removing an um mid-sentence can make two words collide. If the sentence sounds rushed after the cut, put it back.
- Authenticity. A casual vlog with no fillers at all can feel scripted. A few left in sound natural.
A practical target, based on the 2024 study, is to bring filler sounds down to a handful per minute in the final edit, not to remove every one.
Editing decisions, case by case
| Situation | Decision |
|---|---|
| Um or uh at the start of a sentence | Cut. These are the easiest to remove cleanly |
| Um in the middle of a sentence, followed by a pause | Cut both the um and the pause, or the gap stays |
| Restart such as “we, we went” | Cut the first attempt and keep the final take |
| “Like” or “so” as a filler | Cut only if the sentence still sounds natural |
| Um before an answer in an interview | Usually keep |
| Many fillers in one passage | Cut the sounds, then re-record if it still sounds hesitant |
The second row is the one people miss. Removing an um often leaves the silence around it, so the listener hears a gap instead of a filler. Run silence removal after filler removal, or use a tool that does both in one pass. What is dead air covers how long a pause can run before it becomes a problem.
Removing fillers from a recording
A tool that only measures loudness cannot find filler words. An um is as loud as the words around it. You need speech recognition, which is why the free tool on this site removes silence but not fillers.
Premiere Pro
Adobe added filler word detection to text-based editing. According to Adobe’s announcement, it detects sounds such as uh and um and labels them as [filler] in the transcript.
- Transcribe the sequence. Clips transcribed before the feature existed need to be transcribed again to show fillers.
- In the Transcript panel, click the filter icon and choose Filler.
- Click Delete, then choose Extract to close the gaps or Lift to leave them.
- To change what counts as a filler or pause, open the panel’s overflow menu and choose Transcribe view options.
Adobe’s announcement says filler detection works in all 18 languages supported by Speech to Text at the time. It was announced as a beta, and some users in the thread report changes in later builds, so check your version.
Descript
Descript’s help center describes Remove filler words in the AI Tools panel in the right sidebar. You can delete each filler, which removes text and audio, or ignore it, which strikes the word through in the script while removing the audio. On current plans the feature uses AI credits. Descript’s help pages state that filler word detection currently works on English transcripts only.
QuellCut on iPhone and iPad
- Import the video or audio.
- Open Enhancements and choose the language spoken in the recording.
- Turn on Remove filler words. Turn on Remove repetitions to keep only the final take of a repeated sentence.
- Review every cut on the timeline and toggle back any you want to keep.
QuellCut uses Apple speech recognition, on the device where your language supports it, and removes silence in the same pass. The filler words guide has more detail on each method.
Reducing fillers while you speak
Editing fixes a recording. For live talks and calls, a common piece of advice is to replace the um with a silent pause. A short silence also signals that you are thinking, and in a recording it is easy to trim afterwards. Slowing down slightly and knowing your first sentence before you start can help too.
If you record regularly, count your rate again on a later recording with the same method, so the numbers are comparable.
Next steps
If your recording also has long pauses, the free tool on this site can find and remove them in your browser for files up to 3 minutes, without uploading anything. It leaves filler words in place, since it measures loudness only.
For filler words, repetitions and silence in one pass on your phone, QuellCut runs the edit on the device and lets you review each cut before you export. For short-form clips, see Remove silence for Shorts, Reels and TikTok.
Questions
- Is it normal to say um?
- Yes. Research on spontaneous speech treats um and uh as ordinary parts of how people speak, used to signal a delay while planning what to say next. The goal is a low rate, not zero.
- How many filler words per minute is acceptable?
- The authors of a 2024 study in the Journal of Applied Behavior Analysis concluded that five or fewer per minute may be acceptable. At 5 per minute, perceived effectiveness was not harmed in their test.
- Are um and uh worse than like or so?
- In that study, filler sounds had a stronger effect on listener ratings than filler words. Cut um and uh first when you edit.
- Can I remove filler words automatically?
- Yes, with a tool that uses speech recognition, such as Premiere Pro's transcript, Descript or QuellCut. A tool that only measures loudness cannot find them, because an um is as loud as any other word.
- Does Descript remove filler words in other languages?
- According to Descript's help center, its filler word detection works on English transcripts only. Check the current help page, since support can change.
Sources: Laske & DiGennaro Reed (2024), Journal of Applied Behavior Analysis, doi:10.1002/jaba.1093, Clark & Fox Tree (2002), Cognition 84, doi:10.1016/S0010-0277(02)00017-3, ABAI blog: Followup on form and function in verbal behavior, Adobe Community: Filler words in Premiere Pro text-based editing, Descript Help: Remove filler words.