Clarity: the axis with a formula
By Noam Shemla · Data & Method
TL;DR Clarity is the one axis you can partly audit with arithmetic. A formula starts at 100 and subtracts six points for every one percent of your words that is a filler. But the formula does not own the axis. It is one of four things feeding it, alongside a model score, your pace, and your hesitation pauses.
Clarity here means one specific thing: how easy your message was to follow, and in practice, mostly the proportion of your words that carried no content.
Articulation, diction and whether your argument made sense all fall outside it; the last of those is Structure, scored separately. The definition is deliberately narrow, because this is the part of "clarity" a recording can be interrogated for.
Four inputs, one score
Deepgram nova-3, word-level timestamps with per-word confidence
Measured
- Filler rate, % of words
- Hedging rate, % of words (sessions only)
- Words per minute
- Pause locations and lengths
Inferred, not measured
- The Clarity score itself, which the model also assigns directly
Four things move this axis, weighted as follows:
The formula is short enough to print. It starts at 100 and subtracts six points per percentage point of filler, three per percentage point of hedging:
The part where the two paths disagree
Hedging enters that formula at half the weight of filler, but only on live-session transcripts. A recorded practice passes filler only, and its hedging count is never applied.
This matters because of an overlap. Four phrases (I think, I guess, kind of, sort of) appear on both the filler list and the hedging list. On a session transcript they are therefore charged twice, at six points and again at three:
Left out of the peer vote
Clarity is the only axis, along with Intellect, that peer votes cannot touch. Members voting in a Battle or a Round score tonality, body language and narrative, and their overall impression is spread only across the axes an audience can perceive.
That exclusion is the design working as intended. A listener's sense of how articulate someone was is precisely the unmeasured intuition this axis exists to supplement, and folding it back in would be a circle.
What the studies say about filler
The case against filler is weaker than the coaching industry implies, and worth knowing before you spend a month trying to eliminate it.
Disfluency carries information. An uh before a word reliably predicts that the speaker is about to say something hard to retrieve, and listeners exploit that: they recognise the following word faster than they do without the warning (Fox Tree, 2001).1 Listeners also read hesitation as a real signal about how confident a speaker is in their answer, which means removing it removes something honest (Brennan and Williams, 1995).2
The case against hedging is somewhat stronger, and older. Work on courtroom testimony found that speech marked by hedges, hesitations and intensifiers (a cluster the authors called "powerless speech") produced lower ratings of the speaker's credibility and competence than the same content delivered without them (Erickson and colleagues, 1978).3 Later work qualified this too: hedges can help when an audience is already resistant, by making a claim easier to accept.
What survives is modest and it is the version this axis is built on: density is the problem, not presence.
What the filler count gets wrong
It cannot tell a filler from a word. The wider of the two lists includes well, okay, right, yeah, honestly, obviously and I think. Every one of those has legitimate uses. Your filler percentage is an upper bound on the problem.
Two different lists run on the same recording. Both counters run on every measured clip, so this has nothing to do with which path you took. One list of sixteen patterns produces the filler count and the timestamps the coaching text refers to; a second list of twenty-four produces the filler rate that is stored, trended and benchmarked. The wider list adds well, okay, right, yeah, honestly, obviously, anyway and to be honest. So the count and the percentage on the same report are not arithmetically consistent with each other, and the percentage is the more generous of the two.
It has no baseline for you. Three percent may be ordinary for one speaker and a red flag for another. The formula treats them identically.
It rewards silence it cannot see. Replacing an um with a deliberate pause raises this score. Replacing it with a rushed continuation also raises it. The axis cannot distinguish the two, and only one of them is an improvement.
Bringing the filler rate down
The reliable route is substitution. Put a pause where the um was going to go. It costs nothing on any axis and gains you the one thing filler never provides, the appearance of deliberation.
The second route is preparing openings specifically. Filler clusters at the start of sentences and at transitions, because that is where retrieval load peaks. A speaker who knows their first sentence and their transitions cold usually finds the rate falls without ever having worked on it directly.
References
- Fox Tree, J. E. (2001). Listeners’ uses of um and uh in speech comprehension. Memory & Cognition, 29(2), 320–326. https://doi.org/10.3758/bf03194926
- Brennan, S. E., & Williams, M. (1995). The feeling of another’s knowing: prosody and filled pauses as cues to listeners about the metacognitive states of speakers. Journal of Memory and Language, 34(3), 383–398. https://doi.org/10.1006/jmla.1995.1017
- Erickson, B., Lind, E. A., Johnson, B. C., & O’Barr, W. M. (1978). Speech style and impression formation in a court setting: the effects of “powerful” and “powerless” speech. Journal of Experimental Social Psychology, 14(3), 266–279. https://doi.org/10.1016/0022-1031(78)90015-x