English Filler Words, and How to Ask Someone to Repeat

"So what made you want to move here?"

Two seconds of nothing. Then: "The weather."

Now the same answer with one word in front of it. "Um… honestly? The weather. And my sister's here."

The second person sounds like they are thinking. The first sounds like the question did not land. That reading of the silence is mine, not a finding: nobody has measured how listeners interpret an L2 speaker's pause. The research goes only this far. A filler makes your listener attend to what comes next.

Grammarly's page on filler words ranks near the top of Google for this topic, and it has a section headed "Embrace the silence." Its advice: "If you're not sure of what to say next, don't say anything at all… a small silence is often better than a string of filler words." That is written for a native speaker rehearsing a presentation, where a deliberate pause reads as composure. If your English is still assembling itself while you talk, the identical pause reads as lost.

Below, both halves of stalling out loud in English. First: how often fluent speakers actually stall, which filler does which job, which ones survive a job interview, and the hesitation sound your first language keeps smuggling into your English. Then the half nobody teaches, which is what to say when the sentence you missed was theirs.

In a lab corpus, English speakers stalled six times every hundred words

Heather Bortfeld and four colleagues recorded 48 pairs of people doing a picture-matching task and counted every stumble in roughly 192,000 words. Speakers produced 5.97 disfluencies every 100 words, of which 2.56 were fillers and the rest were repeated words and false starts. All 96 speakers were adults, college-educated, recruited around Stanford, and none of them was being tested on a foreign language.

The interesting number in that study is not the average. Within each pair, one person had to describe the pictures and the other only had to find them. The describers hit 7.00 disfluencies per 100 words; the finders 4.93. Fillers alone: 3.30 against 1.81. Same people, same room, same language. The one doing the harder planning job stalled more.

That is what a filler measures: planning load. Not vocabulary and not confidence. Sharon Oviatt's numbers, quoted inside the same paper, make the point from the other direction: people talking to each other produced 5.50 to 8.83 disfluencies per 100 words, people delivering a monologue 3.60, and people talking to a machine between 0.78 and 1.87. Nobody stalls when there is nothing to negotiate.

Um and uh are English words, and they do different jobs

Herbert Clark and Jean Fox Tree went through the London-Lund corpus of British conversation, 93 speakers, and found that um and uh behave like vocabulary. Their conclusion: uh and um are conventional English words, planned and produced the way any other word is.

They also do different work. Um was followed by a delay 61% of the time against 29% for uh, and the pauses after it were longer, 0.68 units against 0.25. So um announces a bigger wait than uh does. In the Switchboard corpus of American telephone calls, 2.7 million words, uh appears 67,065 times and um 12,558. Neither is new. Etymonline dates written um to the 1670s and uh to around 1600, used for hesitation by 1926.

The strongest evidence that fillers do something for your listener comes from an experiment where the control condition was a cough. Scott Fraundorf and Duane Watson played 72 people a recorded story and tested what they remembered. Plot points introduced after a filler were recalled with 1.57 times the odds of the fluent version, 95% confidence interval [1.01, 2.46]. Then they spliced in coughs of matched duration. Recall got worse: odds of 0.64, interval [0.44, 0.93]. The delay was identical. The um helped and the cough hurt, so whatever is happening is not about buying time in the abstract. Your listener is using the sound.

The English filler words list, sorted by what each one does

FluentU lists twenty-plus English fillers. Preply lists twelve. Both run as flat, ungrouped lists, which is useless at the moment you actually need one. A filler word is a sound or phrase that holds your turn while you plan the rest of the sentence. Sorted by which part of that job each one does, there are four groups and about a dozen words worth owning, with an example of where each one fits.

Buying time. Um, uh, er, erm, well, so. Cambridge's grammar reference says the main function of well is to show that we are thinking about the question we have been asked, with the example A: How long would it take to drive from Dundee to here? B: Well, let me see. This is the safest word in the whole list, and the one most learners never use.

Softening. Cambridge's hedge list: apparently, arguably, I think, just, kind of, like, maybe, perhaps, presumably, probably, roughly, sort of, surely. "The flight is expensive" is a fact you are asserting. "The flight's kind of expensive" leaves room for someone to disagree with you without a fight. Cambridge notes that sort of is more common in British English and kind of in American English.

Checking. You know, right?, does that make sense? Fox Tree surveyed 105 undergraduates about what these markers mean to them and found that you know was thought to be used in checking for understanding and connecting with listeners. Cambridge's entry agrees: you know checks that we share the same knowledge. Ending a complicated sentence with "does that make sense?" gives the other person a legal opening to say no, which is worth more than it costs you.

Reformulating. I mean, what I mean is, or rather. This is the one that rescues a sentence you have already broken:

A: The rollout is blocked on the vendor. B: Blocked how? A: They haven't… I mean, what I mean is, they've signed, but nobody's given us the keys.

Speaker A abandoned a sentence mid-word and did not restart from the beginning. I mean is the hinge that let them turn.

You know hands the turn toward your listener. I mean keeps it and repairs what you just said. Both get called filler.

Which fillers are safe where

Like is not banned everywhere. It is banned in one place, and knowing which place is the entire skill.

Most people already sort their fillers this way without being taught. In that same survey, 66% of respondents said they use these markers more with friends and less with authority figures. Only 12% claimed the reverse. What the survey does not tell you is which cell each individual word belongs in. Those calls are mine.

A caution about how permanent any of this feels. Merriam-Webster's defence of like points out that a century ago the word under attack was well: Ambrose Bierce, in his 1909 book Write It Right, called it "a mere meaningless prelude to a sentence". The filler this post recommends as universally safe was, within the lifetime of the language, the one people complained about.

I cannot tell you that like has cost anyone a job, because no study I could find tests a single filler against a hiring decision. What the evidence supports is narrower and still useful: like sits on the hedge list, native speakers already suppress it in front of authority figures, and an interview is the highest-stakes place you will ever be judged on twenty minutes of speech.

The interview-safe version of stalling sounds like this:

Interviewer: Why are you leaving your current role? Candidate: Well, I've learned a lot there, but I'm not being stretched any more.

One word, and the candidate got a beat to think. One warning on it from Cambridge: well on its own with rising intonation, as a question, can sound very direct. "Well, I've learned a lot" is safe. "Well?" is not.

The hesitation sound that gives away your first language

Mandarin's most common filled pause is 那个 (nàge, also pronounced nèige), literally "that". Yao Zhao and Dan Jurafsky measured it in a corpus of Mandarin telephone conversations at 4.51 occurrences per 1,000 words, nearly twice the rate of Mandarin's uh at 2.55. For comparison, they put English um at 7.15 and uh at 7.10 per 1,000 words in the CallHome corpus. Both languages stall constantly. They just reach for different sounds.

That sound follows people across languages. Marie Böttcher and Margaret Zellers analysed over 9,000 fillers from Russian, English and German speakers and found the majority language reshaping filler choice in the heritage language, though bilinguals still kept the two sets partly apart. Learning the English set is a thing you do on purpose.

Two of these deserve a note. Japanese speakers have more than one option and they are not interchangeable: Takanori Kawada, working with the Corpus of Spontaneous Japanese, describes あの ano as oriented toward the listener while ええっと eeto is the sound of the speaker computing something internally. Spanish speakers reaching for este are reaching for something dominant rather than marginal. Lamar Graham counted 9,713 hesitation markers in a 2000s corpus of Mexico City speech and found 75.7% of them were este rather than eh, rising to 82.3% among younger speakers. Russian speakers reach for ну, roughly "well", which does several jobs at once in the way that Russian discourse particles generally do. Korean has 저기 jeogi and 음 eum.

那个 needs its own paragraph. English Wiktionary carries a usage note warning that English speakers may mishear nèige as a racial slur, and in 2020 a USC business professor was removed from a course after saying it in class. The class was about filler words. Whatever you think that case should have concluded, the practical lesson for a Mandarin speaker talking to English listeners is small and clear: this is the filler to swap out first.

Ask on the first sentence, while asking is still cheap

You nod. You smile. You say "yes, yes" and hope the next sentence brings you back. It does not, because the next sentence assumes you got the last one, and now you are two behind instead of one. By the fourth sentence, asking has become embarrassing in a way it was not at the start, so you keep nodding. The whole conversation is spent, and neither of you knows it yet.

Asking is not the unusual move here. Mark Dingemanse and eleven co-authors went through 48.5 hours of recorded conversation in twelve languages from eight language families and found 2,053 moments where someone stopped to sort out a problem in what had just been said. That is a repair initiation about once every 1.4 minutes. On average, five minutes did not go by in any of the twelve languages without one. English speakers interrupt each other for clarification constantly. You are not adding friction to a smooth conversation; you are doing the thing conversations are made of.

If the sentence dissolved on you rather than being too quiet, the cause is usually that English ran the words together, and connected speech has patterns you can learn to hear. If you heard every word and still lost the sentence, the cause is probably a corporate jargon phrase doing social work instead of describing anything.

Three ways to ask someone to repeat in English, and which one to use first

Dingemanse's team sorted every one of those 2,053 repairs into three types, which between them covered the vast majority of cases in every language, 92% on average.

Open. "Sorry?" "Huh?" "What?" You signal that something went wrong without saying what. Average length across languages: 3.7 characters.

Restricted request. You name the missing piece. Cambridge's example: A: His name is Thokosani. B: His name is what? Average length 10.4 characters.

Restricted offer. You hand over a guess and ask them to confirm it. Cambridge again: A: Did you hear Pete's giving up his job. B: Pete's giving up his job? Average length 13.0.

The lazy move is "Huh?", because it makes the other person redo everything while you do nothing. The researchers found that people mostly do not take it. They pick the most specific form they can manage, which loads more of the work onto themselves, and the paper calls this an element of altruism in how people initiate repair. Being specific is the polite option.

Open forms have their moment. Where there was actual noise, or someone talking over the speaker, the estimated probability of an open initiator ran around 76 to 78%, against 39% in clean conditions. If a truck went past, "Sorry?" is exactly right. If you heard every sound and one word did not land, "Sorry, whose office?" is better.

Skip the museum pieces. 7ESL's list of forty ways to ask still offers "Would you be so kind as to repeat that?" and "I beg your pardon, would you mind repeating that?" Nobody under seventy says either in an office. "Sorry, I didn't catch that" covers almost every situation you will meet, and "Sorry?" covers the rest.

What to say on the second and third ask

They repeated it. You still did not get it. This is the moment that no phrase list prepares you for, and it has a documented shape: in Dingemanse's data, once an open initiator like "Huh?" had already failed, the chance that the next attempt was also open fell to 17.4%. Across twelve unrelated languages, people almost never ask the same way twice. They get specific.

So change the variable. Repetition at the same speed rarely fixes anything, because the problem was almost never loudness.

A: Can you send it to Aoife on the fourth floor? B: Sorry, I didn't catch that. A: [louder] Can you send it to Aoife on the fourth floor? B: Sorry, the name? Could you spell it? A: A-O-I-F-E. Ee-fa. B: Got it. Aoife, fourth floor.

Four moves worth having ready when the repeat fails:

• "Sorry, could you say that a different way?"
• "Could you spell that?"
• "Could you slow down a little?"
• "So you want it by Friday, is that right?"

The last one is the restricted offer, and it is the strongest of the four when you are 80% sure. You hand over your best guess and they correct one word instead of rebuilding the sentence.

What to do this week

Record two minutes of yourself talking about your day, with no script. Play it back and count two things: the silences longer than about a second, and the number of times a sound from your first language came out instead of um. Those are the two places to work, and they are separate jobs. The first one needs well and so in your mouth until they arrive without being summoned. The second one needs you to hear este or 那个 or あの on the recording, which is the hard part, because you have never once heard yourself say it.