AI can't stop you saying "um" mid-sentence, but it is very good at the part most people skip: counting. Record a few answers, get a word-for-word transcript, and ask a chatbot to count and sort every filler, show where in your answers they cluster and suggest what goes in each spot instead. The fix itself is physical, a short silent pause where the filler was, and only out-loud repetition makes it stick.
Below: the measure-fix-remeasure workflow, why some transcripts quietly delete your ums, five prompts, a before-and-after answer and a 10-day plan. For the speaking technique in depth, see our guide on how to stop saying um in interviews.
What AI can and can't do about filler words
| Part of the job | AI helps by | You still have to |
|---|---|---|
| Counting | Counting and sorting fillers in a transcript in seconds, including phrases such as "you know" and "kind of" | Make sure the transcript kept the ums (see the next section) |
| Finding patterns | Showing whether fillers cluster at the start of answers, between story parts or before numbers | Listen for the same pattern in your next recording |
| Replacements | Suggesting short bridge phrases and where to pause | Practise the pause until it feels normal |
| Hedges | Flagging "I think", "maybe" and "sort of" in front of your results | Decide whether the claim is solid; if it is, say it plainly |
| The habit itself | Nothing while you're speaking | Repeat short, timed answers until the pause replaces the filler |
First, get a transcript that keeps your fillers
This step decides whether your count means anything. Speech-to-text services differ: according to their developer documentation, some remove "um" and "uh" by default and others keep them, as our page on whether filler words hurt AI interview scores explains with sources. Phrases such as "like", "you know" and "basically" are real words, so they almost always survive.
- Spot-check by ear. Play the first 30 seconds of a recording and tally the ums and uhs on paper. If the transcript shows fewer, it has been cleaned: count sound fillers by ear and let the AI count the rest.
- Note the speaking time. Write down how long each answer took. You need it to turn a count into a rate per minute.
- Use real interview answers. Five spoken answers give a fairer count than reading a paragraph aloud, because fillers rise when you're composing on the fly.
Our guide on how to record yourself for interview practice lists free recorders and a simple setup.
The measure, fix, re-measure workflow
- Record five answers to common questions, timing each one.
- Get a transcript and spot-check it by ear.
- Run the count prompt (prompt 1) for totals by type and a rate per minute: fillers divided by minutes spoken.
- Run the map prompt (prompt 2) to see where the fillers sit.
- Pick one spot to fix first, usually the warm-up "so" at the start or the "um" before a number.
- Drill that spot for two or three days with its replacement: a planned first sentence, a pause or a bridge phrase.
- Re-record the same five questions, re-count and compare (prompt 5).
Your own first count is the number to beat. Aim for a steady fall, not zero: a delivery scrubbed of every natural pause can sound read, which is its own problem (see how not to sound rehearsed in an interview).
Where fillers cluster, and what to put there instead
| Spot in the answer | Typical filler | Replacement to drill |
|---|---|---|
| The first word | "So...", "Um, well..." | Plan your first sentence before you speak and open with the headline |
| Between parts of a story | "And then, like...", "So basically..." | A one-second pause, or a signpost such as "The first thing I did was..." |
| Before a number or a name | "It was, uh, about..." | Pause, then say it; rehearse the two or three numbers in each story |
| Before a claim | "I kind of led...", "I guess it helped" | Say what you did plainly, or say honestly what you don't know |
| The end | "...so yeah", "...and stuff like that" | A closing line that restates the result, then stop talking |
| A restart | "We, I mean I, well the team..." | Pause, then start the sentence once, cleanly |
Five prompts for counting and fixing fillers
1. Count and sort
Here is a word-for-word transcript of five spoken interview answers, with speaking times: [times]. Count every filler and sort them into: sound fillers (um, uh, er), filler words (like, you know, basically, actually, right), hedges (kind of, sort of, I guess, I think, maybe, probably), false starts and repeated words, and trailing endings (so yeah, and stuff). Only count "like" when it isn't a verb or a comparison. Give a table with counts per answer and a rate per minute. Transcript: [paste]
Why it works: sorting by type tells you which fix to start with; a single total doesn't.
2. Map the spots
Using the same transcript, show where the fillers happen: at the start of an answer, between parts of a story, before a number or name, before a claim, or at the end. Quote three examples of my most common spot. Don't rewrite anything.
Why it works: most people have one or two trigger spots, and fixing a spot is easier than fixing a word.
3. Bridges that sound natural
My most common filler spot is [spot], where I say [filler]. Suggest four short bridge phrases or pause cues for a spoken interview answer, five words or fewer each, that sound natural rather than scripted. Then give me a five-minute drill to practise them out loud.
Why it works: you need something to put in the gap, or the old filler comes back.
4. Hedge audit
List every place in this transcript where I softened an action or a result with a hedge (kind of, sort of, I think, maybe, probably). For each, say whether the sentence would be stronger said plainly or whether the hedge is honest because the fact is uncertain. Transcript: [paste]
Why it works: hedges in front of your results do the most damage, but some are accurate, and you should keep those.
5. Before and after
Here are two transcripts of the same five answers, before and after a week of practice, with speaking times. Compare the filler rate per minute by type and tell me which type fell least. Before: [paste]. After: [paste]
Why it works: a measured change keeps you practising; a feeling of improvement often isn't reliable.
Worked example: one answer before and after
Question: "What's your biggest professional achievement?" An example we wrote; the details are illustrative.
Before (transcript, about 35 seconds)
"Um, so, I think my biggest achievement was probably, like, the warehouse layout project. So basically we, I mean I, was asked to, um, look at why picking was slow, and I kind of mapped out, you know, where everything was, and, uh, we moved the fastest-selling items closer to packing. And it, like, helped a lot. I think pick times went down by, um, about a fifth? So yeah."
What the AI found (prompts 1 and 2)
"16 fillers: sound fillers 4 (um ×3, uh ×1), filler words 6 (like ×2, so ×2 at the start of a sentence, basically, you know), hedges 4 (I think ×2, probably, kind of), false starts 1 (we, I mean I), trailing 1 (so yeah). Most sit at the opening and in front of the result, where 'I think' and 'um' weaken your best number."
After (re-take on day 6)
"My biggest achievement was speeding up picking in our warehouse. I was asked to find out why picking was slow. I mapped where every item sat, and we moved the fastest-selling items next to packing. Pick times went down by about a fifth."
Same facts, nothing added. The candidate planned the first sentence, paused before the number instead of saying "um", and ended on the result instead of "so yeah". A short answer like this one can come out clean; in longer answers, a falling count is the realistic goal.
Check the AI's count and advice
- Miscounts. "Like" as a verb ("I'd like to"), "so" meaning "therefore" and "right" as a direction aren't fillers. If a number looks off, ask the AI to quote every instance it counted.
- Cleaned transcripts. A transcript that dropped the ums undercounts. Combine the AI's count with your tally by ear.
- Over-correction. Advice that pushes you towards a flawless, filler-free script will make you sound recited. Pauses are fine; scripts aren't.
- Honest hedges. Keep "I think" when you genuinely don't know. Interviewers value accuracy over certainty.
- Privacy. Remove client names, colleagues' names and confidential figures before pasting a transcript.
- Not a clinical tool. If a stammer or another speech condition is involved, a speech-language professional is the right person to see. These drills are about habit.
Our guide on how to use AI for interview prep shows where delivery work fits among the other prep tasks.
A 10-day plan to cut filler words
| Days | Drill | Measure |
|---|---|---|
| 1 | Record five answers; run prompts 1 and 2 | Your starting rate per minute and your top spot |
| 2–3 | Fix the top spot only: plan your first sentence, or pause before numbers | Re-record two answers each day |
| 4–5 | Hedge audit (prompt 4); say every result plainly | Hedges in front of results |
| 6 | Re-record the original five; re-count (prompt 5) | Rate compared with day 1 |
| 7–8 | Your second spot from the map, with bridges from prompt 3 | Two answers a day |
| 9 | A full AI mock interview round with questions you haven't seen | Fillers in the transcript or recording |
| 10 | Re-record the original five | Final rate; keep one counted session a week until the interview |
If the rate isn't falling after a few days, change the drill rather than repeating it: slow down a little, plan your first sentence, or shorten the answer so there's less to compose on the fly.
Where CoPilot Interview fits
CoPilot Interview is an AI interview assistant. Its AI mock interview reads each question aloud and scores your spoken answer out of 10, which gives you a steady supply of unrehearsed questions to re-count against. The free plan gives 15 minutes a day of live help in sessions of up to 15 minutes, with AI mock interviews included, so one counted answer a day costs nothing. Paid plans are Standard at $24.99 a month and Pro at $39.99 a month.
In real interviews, the desktop app's live help shows captions of the interviewer's question and a structured answer in about two seconds. If you use it, follow your employer's rules. Online assessments, proctored tests and AI-led rounds typically forbid outside help, so prepare for those instead.
Disclosure: CoPilot Interview is our product, so weigh what we say about it accordingly.
Frequently asked questions
Can AI count my filler words?
Yes. Paste a word-for-word transcript of your answers into any chatbot and ask it to count and sort the fillers, with a rate per minute. Spot-check by ear first, because some transcription tools remove "um" and "uh" by default.
How many filler words are too many in an interview?
There's no official limit. Aim for a falling rate from your own starting count and for clean sentences where they matter most: your opening, your numbers and your result.
Do filler words lower AI interview scores?
A few ums usually matter less than people fear, since AI scoring tends to work from a transcript and some services strip sound fillers. Filler phrases, hedges and false starts survive transcription and can blur your evidence.
Should I pause instead of saying um?
Yes. A short silent pause in place of a filler sounds thoughtful to a listener even when it feels long to you. Close your mouth at the end of a sentence, breathe and start the next one.
How long does it take to cut filler words?
It varies from person to person. Re-count every few days; if the rate isn't falling, change the drill (slow down, plan your first sentence, pause before numbers) instead of simply repeating it.
Rehearse it out loud before the real one
The AI mock interview builds questions from your resume and the job description, reads them aloud and scores each spoken answer out of 10. Included on the free plan.
Start a mock interview