Filler words

Why does a speaker who pauses to say \"um\" get understood better than one who never hesitates?

Every course in public speaking gives the same instruction: cut the ums. And yet the seamless speaker — the one who never stumbles, never stalls, never audibly searches — is often the harder one to follow. Experimentally, a word arriving just after uh is recognised faster than the same word arriving smoothly, and stretches of speech containing filled pauses are remembered better afterwards.

That is odd enough to be worth pulling apart. The speaker's reason for saying um is obvious: they need a moment. But the speaker's need does not explain the listener's benefit. So what work is the um doing on the receiving end?

The first candidate is simply time. The filled pause takes a few hundred milliseconds, and the listener, who is also running behind, gets to use them. Plausible — and testable, because it predicts that any delay of the same length should do the same job. A silent gap should be as good as uh.

It is not. A silence is ambiguous in a way a filled pause is not: it might mean the speaker is thinking, or that they have finished, and a listener who reads it the second way takes the floor. Uh resolves that. It says the turn is still claimed and something is still coming. So the filled pause is not time — it is a message about the time, and that difference is the whole phenomenon.

Once you see it as a message, the next question is why anyone should believe it. The answer is that it is not a claim at all but a symptom, which is exactly what makes it trustworthy. Speakers hesitate more before low-frequency words, before proper names, before a thing introduced for the first time rather than one already under discussion. The hesitation is caused by the difficulty, so it cannot easily be uncoupled from it — not a signal chosen, but a leak that reliably tracks what is coming.

Listeners exploit that. Eye-tracking work by Arnold and colleagues found that on hearing a drawled article followed by a filled pause, listeners shifted their gaze toward the object that had not been mentioned before — using the hesitation to predict novelty before the noun arrived. The tell that this is learned rather than hardwired is what happens when the contingency is broken: in a follow-up, listeners told the speaker had a condition affecting object naming stopped using the hesitation that way. That study is the softest evidence in this piece and worth checking directly, but the logic it supports is solid: a proxy is only informative while its cause is intact.

There is a second job too, about repair rather than prediction. Speech is planned while it is spoken and it goes wrong constantly. When a speaker breaks off — the yell-, uh, the purple square — the break marks where the error was, so the listener knows which stretch to discard. Delivering the correction smoothly hides that boundary and leaves the listener to find it.

So the fluent speaker has not removed noise. They have removed a running commentary on their own planning: where the hard parts are, where a new thing is entering, where something just said should be thrown away.

Three qualifications keep this honest.

The effects are real but small, and they are effects of ordinary rates of disfluency. A speaker who stalls in every clause, or repairs the same phrase four times, does register as struggling, and listeners judge accordingly. Nothing here recommends more ums; it only says the ones that occur naturally are not waste.

The claim that um and uh differ systematically — that um precedes the longer delay — was argued by Clark and Fox Tree and is genuinely contested, with later corpus work finding the distinction weaker and less consistent across speakers and languages than first proposed. Treat the two as broadly interchangeable unless you have data on the particular speaker.

And the cue cuts both ways. Disfluency rises with unfamiliarity, stress, and speaking in a second language, so the same hesitation that helps a listener parse the sentence also invites a judgement about the person producing it. This sits beside the collection's piece on transcribed speech, which asks the opposite question — why a verbatim record of that same behaviour looks broken on the page. There the disfluency is what makes the transcript unreadable; here it is what made the speech easy to hear. Both are true, because the page has no channel for what the ear was using it for.

• Why scripted speech read aloud is harder to follow than the same content spoken from notes.
• How speakers of different languages fill their pauses, and whether listeners can use a foreign filler at all.

Every term the collection defines is gathered in the glossary.