How many filler words per minute is normal?

Count every um, uh, false start and mid-sentence repeat, and spontaneous speech runs to roughly 6 disfluencies per 100 words. Bortfeld and colleagues measured 5.97 per 100 words on a referential-communication task; the Switchboard telephone corpus that Elizabeth Shriberg analysed came in at 6.4. Fillers alone – just the um and uh – are a smaller slice, about 2 to 3 per 100 words.

That word roughly is doing real work. There is no single universal filler rate, because it moves with the situation. The same Switchboard study found people talking to an automated travel system produced only 0.4 disfluencies per 100 words, against 6.4 talking to another person and 13.3 in a corpus of phone calls where people planned real air travel with an agent. How much planning the speaker is doing, how familiar the material is, how formal the setting – all of it shifts the number. So the honest headline is a band, not a point.

It helps to separate two things people lump together. A filler is a filled pause: um, uh, er. A disfluency is the wider family – fillers plus repeats ("the the") plus restarts, where you abandon a phrase and begin again. When a headline says six per 100 words, it usually means the whole family.

Bortfeld's corpus breaks it down cleanly. Of the 5.97 disfluencies per 100 words, fillers accounted for 2.56, repeats for 1.47, and restarts for 1.94. So if you only care about um and uh – the sounds people actually notice and want stripped from a transcript – the number to hold onto is closer to two-and-a-half per 100 words than to six.

Research counts disfluencies per hundred words because words, not seconds, are what a transcript hands you. To get a per-minute figure you need a speaking rate. The National Center for Voice and Speech puts conversational English at around 150 words a minute, with everyday talk usually falling between 120 and 150.

Do the arithmetic at 150 words a minute and the picture sharpens. About 2.5 fillers per 100 words works out to roughly 3 to 4 um or uh every minute – one every 15 to 20 seconds. Count the full disfluency family at 6 per 100 words and you are near 9 a minute, one every 6 or 7 seconds.

Two cautions on that conversion. It is arithmetic, not a stopwatch measurement – no study clocked a clean "one filler every N seconds" as a constant. And it scales with pace: slow down and the interval stretches, speed up and it shrinks. Treat it as an order of magnitude, not a metronome. If you want the full arithmetic of words to minutes to hours, we ran it in how many words are in an hour of audio.

When Herbert Clark and Jean Fox Tree counted fillers across the full 2.7-million-word Switchboard corpus, they found 79,623 of them – and uh outnumbered um roughly five to one: 67,065 uhs (84 percent) to 12,558 ums (16 percent), a little under 3 fillers per 100 words overall.

But that 84 percent is not a law of nature; it is a fact about telephone speech. The same authors found the share that is uh swings by setting: 84 percent on the phone, 76 percent on answering machines, 54 percent among British academics talking face to face, and just 44 percent in narratives. Register moves the mix.

There is also a signal buried in the choice. Clark and Fox Tree argued that um announces a longer coming delay than uh. A speaker's tendency to pause was predicted far better by how often they said um (a 0.48 correlation, about a quarter of the variance) than by how often they said uh (0.25, which explained almost none of it). Um is the sound of a bigger hitch.

Three things reliably move a person's filler rate, and none of them is a character flaw.

Planning load. In Bortfeld's task, the person doing the describing – the director – was markedly more disfluent than the person listening and matching: 7.00 versus 4.93 disfluencies per 100 words, and on fillers alone 3.30 versus 1.81. Harder cognitive work, more ums.

Age. Older speakers were modestly more disfluent, about 6.65 per 100 words, consistent with word retrieval getting a little harder with age.

Gender – but read this one carefully, because it is mostly about which filler, not how many. On Bortfeld's demanding referential task men were actually more disfluent overall (6.80 versus 5.12 per 100 words). Separately, sociolinguists including Wieling, Grieve and colleagues find that women and younger speakers lean toward um while men and older speakers lean toward uh, and that um has been rising over time, led by women and educated speakers. These are different measurements – overall rate, versus the um-to-uh ratio – so they do not contradict each other.

Filled pauses are only part of what people mean by filler. Discourse markers – like, you know, I mean – do similar work and are far more visible in a transcript. In a study by Laserna, Seih and Pennebaker that recorded 263 people going about their days, like was the single most common of them, reported at 1.13 percent of all words spoken, ahead of you know (0.18 percent) and I mean (0.12 percent).

The same work reported that the rate of filled pauses themselves was roughly even across gender, while these discourse markers skewed toward women and younger speakers – another reminder that who says filler words more depends entirely on which words you are counting.

It is tempting to read all this as a catalogue of errors. It is not. Filled pauses are a normal, near-universal feature of unscripted speech, and they carry information: as Clark and Fox Tree showed, um and uh tell a listener that the speaker is thinking and roughly how long the wait will be. People are dramatically more fluent when the words are already planned – which is exactly why talking to a machine produced 0.4 disfluencies per 100 words against 6.4 for real conversation. The ums are the sound of a mind composing in real time.

Which is worth remembering before you judge a transcript. A recording full of um and uh is not a sign of a poor speaker. It is a sign of a human one.

All of this lands on a single practical decision every transcript forces: keep the fillers, or cut them?

Roughly one word in twenty a person says is an um, a false start, or a repeat. A verbatim transcript keeps every one of them; a clean-verbatim transcript strips them so the sentences read the way the speaker meant them. Neither is more accurate – one is more literal, the other more readable. A qualitative researcher coding hesitation wants the ums left in. A podcaster pulling show notes wants them gone. Same audio, opposite right answer. Our guide to verbatim versus clean verbatim walks through when each one is correct.

This is where a tool's job is to stay out of the way. Pepys is an audio-to-text tool that transcribes the raw words first – every um and uh included – so the choice of whether to keep or cut them stays yours, not something decided for you by a model that silently smoothed the transcript before you saw it. Want the cleaned version? Ask for a summary or clean copy after the fact. Want the literal record? It is already there.

And if fillers are wrecking readability because the audio itself is rough – heavy accents, people talking over each other – most of the fix is upstream, before you ever hit transcribe. We cover the practical version in how to improve transcription accuracy, and the specific case of overlapping voices in how well AI handles crosstalk.

Last verified: 2026-07-13.