Why we don’t count your filler words

Everyone says “um”. The commonly cited base rate is around six disfluencies per hundred words in ordinary speechFox Tree 1995 · weak evidence
The ~6/100 figure is a background estimate in a comprehension study — and it excludes silent pauses. — filler words are the normal texture of unscripted talk, not a defect. Experimental interview evidence isolating filler frequency as a cost is thin. We audited it looking for a reason to coach it, and did not find one.

What the evidence actually penalises

The well-supported delivery variable is delay. In recruiter-rated real interviews, pauses before answering averaged under two seconds for the first question and grew with each one — and longer pre-answer delays both degraded answer quality and independently reduced hireabilityBrosy, Bangerter & Mayor 2016 · moderate evidence
First pause averaged 1.88s; delay hurt ratings directly and pushed answers out of story form.. The harm lives in the multi-second silence before you begin — not in mid-answer pausing, which interview research has barely studied, and not in filler count.

Vocal delivery as a whole does carry modest, real signal: a composite of pitch variability, rate, pauses and amplitude predicted supervisor-rated job performance at r = .18–.20DeGroot & Motowidlo 1999 · moderate evidence
Vocal-cue composite from videotaped interviews correlated with later job performance in two manager samples. — small effects, and notably predictive of the job itself, not just the interview rating.

The advice we refuse to give

“Speak faster to sound competent” is extrapolated from persuasion research — a 1975 study of computer-resynthesised voices, which measured competence and benevolence, not interviews and not credibilitySmith et al. 1975 · weak evidence
54 resynthesised voices from 6 speakers; benevolence peaked at a normal rate. Not an interview study.. “Mirror the interviewer” fares no better: the payoff to ingratiation depends on who is rating you — self-presentation styles that pay off in the Western interview context are, in the researchers’ own words, highly selectivePaulhus et al. 2013 · moderate evidence
222 raters of applicant videos; the advantage traced to self-praise and active ingratiation was culture-bound., and impression management correlates far more strongly with interview ratings (r = .47) than with actual job performance (r = .15)Barrick, Shaffer & DeGrassi 2009 · strong evidence
IM works on interviewers far better than it predicts work..

And as scoring policy, never as a measured claim: we do not let accent, fluency or hesitation count feed a competence score. Second-language speakers hesitate more because word retrieval costs more, not because they know less — coaching them to “cut every um” trains anxiety, not articulation. That policy is written into the same rubric this product scores with.