Do filler words matter in AI interviews? A few "ums" usually matter less than you fear, because AI scoring tends to work from a transcript of your words, and several speech-to-text services strip "um" and "uh" from transcripts by default. What does matter is the clutter that survives transcription: filler phrases like "you know" and "basically", false starts, and hedges that bury your point. Those make your evidence harder to find and your answer harder to rate.
HireVue is one vendor that documents its approach. Its 2026 explainability statement says its AI first transcribes your spoken answer with a speech-to-text service, relies only on what you say, and does not analyze tone of voice (as of October 2026). Other vendors may work differently, and assessments that test a spoken language may rate fluency directly, so read your invitation. This page explains what happens to fillers, when they matter and how to cut them without sounding scripted.
What happens to filler words in an AI-scored interview
Your words pass through two steps before anyone rates them: transcription, then scoring.
Step 1: speech-to-text. The transcription service decides whether fillers appear in the text, and providers differ. Here is what three providers' developer documentation says about their default settings, as of October 2026:
| Speech-to-text service | "Um" and "uh" by default | Setting a developer can change |
|---|---|---|
| Deepgram | Stripped out of the transcript | Turn on filler_words to keep them |
| AssemblyAI | Removed before the transcript is returned | Set disfluencies to true to keep them |
| Rev.ai (streaming) | Kept; it defines disfluencies as "ums" and "uhs" | Turn on remove_disfluencies to drop them |
HireVue names Rev.ai as its transcription supplier, but how any vendor configures these settings is rarely public. You can't know whether your "ums" will appear, so don't build your preparation around them.
Step 2: scoring. The transcript is rated against the competencies the employer chose. Fillers that were stripped never reach the scorer. Filler phrases such as "like", "you know", "sort of" and "I guess" are real words, so they are transcribed like any other word. For how this step works, see how AI evaluates interview transcripts.
Step 3, sometimes: a person watches. Recruiters may review the recording itself, and the next interviewer may read your transcript or summary. Human listeners hear delivery, and Mercor's AI interview guide asks candidates to "Articulate clearly and succinctly" (as of October 2026).
Do filler words matter in AI interviews? When they do
- When they hedge your evidence. "We, like, maybe cut it by 20% or something" reads as uncertain in a transcript. The number is there, but the claim is weak.
- When false starts blur who did what. "We, I mean I, well the team..." makes it hard to credit you for the action.
- When they eat the clock. On a timed one-way video, ten seconds of throat-clearing can push your result past the limit.
- When the test is about speaking. If your invitation says the interview assesses spoken language, fluency may be rated as a skill in its own right.
- When a person reviews the video. A recruiter watching the recording hears everything the transcript dropped.
When they don't: an occasional "um" while you pick a word, or a natural "so" at the start of an answer. Nobody needs zero fillers. The goal is clean evidence sentences.
Filler types and how they show up in a transcript
| Filler | What the transcript usually shows | Does it hurt? |
|---|---|---|
| "Um", "uh" | Often removed, sometimes kept, depending on the service | Rarely on its own |
| "Like", "you know", "basically", "kind of", "sort of" | Kept as words | Yes, when it softens your evidence |
| False starts ("We, I mean I...") | Kept, often as fragments | Yes, especially when it blurs "I" and "we" |
| "I think", "maybe", "probably" before a result | Kept | Yes: the result sounds uncertain |
| Repeating the question back | Kept | Mildly: fine once, costly if long |
| Trailing "so, yeah" | Kept | Mildly: replace it with a closing line |
Example: the same answer with and without clutter
Question: "Tell me about a time you saved your team money." Both versions are examples we wrote; the figures are illustrative.
Before (as a transcript might show it): "So, basically, I was kind of like the person who, you know, handled the vendor stuff? And we, I mean I, sort of renegotiated the contract, and I think we saved, like, maybe 15% or something. So yeah."
After: "I managed our three software vendors. When our largest contract came up for renewal, I compared six months of usage data with what we were paying and found 40 seats nobody had used. I renegotiated the contract down to actual usage, which cut the annual cost by about 15%."
Read the "before" version as a scorer would. It's unclear what your role was, whether you or the team renegotiated, and whether the saving was real. The "after" version has the same facts, but every sentence can be quoted as evidence. Notice it isn't robotic: it simply says each fact once, plainly.
How to reduce filler words without sounding robotic
- Know your first and last sentences cold. Fillers cluster where you're unsure what comes next. If you know how you start and where you end, the middle takes care of itself.
- Swap the filler for a short pause. A half-second of silence reads as a clean sentence break. Keep it short; in a conversational AI interview, a long silence can end your turn (see how AI interviewers handle pauses).
- Replace hedges with honest precision. "I think we maybe saved 15%" becomes "We saved about 15%." "About" is honest; "I think maybe" sounds unsure.
- Use signposts instead of "so". "First... Then... The result was..." gives structure that a transcript can follow.
- Prepare bullet notes, not scripts. Read-out answers tend to sound flat, and many scorecard tools flag answers that sound read or AI-written. Our guide to sounding natural in an AI interview covers the rehearsal method.
- Slow down slightly. Speaking a little slower gives you time to choose words, which is often where fillers creep in.
Drills to cut filler words in a week
- Filler audit (day 1). Record two answers on your phone and run them through your phone's built-in transcription, if it has one, or any free transcription app. Circle every filler and hedge. This is your own baseline, not a target.
- Pause swap (days 2–3). Re-answer the same questions. Each time you feel a filler coming, close your mouth for half a second instead.
- Headline and result (days 3–4). For each of your six to eight stories, say only the first sentence and the result sentence until both come out clean every time.
- Hedge hunt (day 5). Rewrite each result without "I think", "maybe" or "kind of". If you don't know the exact number, use "about" and a figure you can defend.
- Scored rounds (days 6–7). Our free AI interview assistant includes a Mock Interview mode that reads each question aloud, listens to your spoken answer and scores it out of 10 with what to fix and a model answer. The free plan includes 15 minutes every day. For a quick browser round, the free AI interview demo has 145 real interview questions across 10 tracks, graded by AI.
If you want delivery metrics such as filler counts and pace, our guide to what AI can measure about delivery explains what practice tools track. If rambling is the bigger problem, start with how to stop rambling in interviews, and if you interview in a second language, see AI interview tips for non-native English speakers.
Take the assessment yourself and follow the employer's rules — use CoPilot Interview to practice beforehand.
Hear how your answers really sound
Answer questions out loud and get each one scored out of 10 with what to fix, so your evidence comes through cleanly on the day.
Try a free AI mock interviewFAQ
Do AI interviewers count filler words?
Some systems might, but vendors that score from the transcript rate what you said against the role's competencies, and several speech-to-text services remove "um" and "uh" by default. Practice tools often count fillers, which is useful feedback but not the same as an employer's score.
Will saying "um" make me fail an AI interview?
Not on its own, in a transcript-scored interview. Several common transcription services drop "um" and "uh" before anything is scored. What costs you is vague or hedged evidence: no clear role, no specific actions, no result.
Is it better to pause than to say "um"?
A short pause is cleaner in a transcript. Keep it to a moment, though: in a conversational AI interview, a long silence can be read as the end of your answer.
Do filler words matter more for non-native speakers?
No rule says so. Clear, complete sentences help any speech-to-text system, so focus on pace and finishing your sentences rather than on your accent. If the assessment tests spoken English itself, fluency may be rated, so read the invitation.
How many filler words is too many?
There's no published threshold. A practical test: read the transcript of your answer and check whether every sentence that carries evidence can be quoted cleanly. If fillers break up your role, actions or result, cut them.
Should I memorize my answers to avoid filler words?
No. Memorized scripts tend to sound read, and many scorecard tools flag answers that sound read or AI-written. Memorize your first sentence and your result, keep bullet notes for the middle, and practice out loud.