What is filler rate?
By Noam Shemla · Measurements
Filler rate is the share of your words that carried no content: um, uh, like, you know. We report it as a percentage of total words, not as a count per minute, so it does not move when you speak faster.
The short version Filler rate is fillers ÷ total words × 100. It costs six clarity points per percentage point. Two different word lists produce two different filler figures inside the same report, and the one shown to you is the wider, so treat it as an upper bound.
Counting the fillers
The transcription service is asked to keep filler sounds in its word list rather than drop them, which it would otherwise do. Every transcribed word is then stripped to its letters and checked against a list: first against the single words, then, if it is not one of them, against the two-word phrases. Kind of is therefore counted once, as a phrase, because neither kind nor of is on the single-word list.
Deepgram nova-3, word-level timestamps with per-word confidence, with filler-word detection enabled
Measured
- Filler count
- The timestamp of every filler
- Filler rate as a percentage of words
Inferred, not measured
- Nothing. This parameter is counted by rule, with no judgement involved
The timed list exists for six languages: English, Spanish, French, German, Portuguese and Hebrew, chosen by the language the transcription service detects. A language outside those six still falls back to English. There is a catch, covered below, that limits how well some of the newer lists can match.
The timestamps matter more than the count. A speaker with four fillers evenly spread has a different problem from one with four fillers in the same sentence. The first is a habit; the second is a moment where they lost the thread.
The two lists
This is the one place where the product disagrees with itself, and it is better to say so than to let you find it.
The narrow list is the timed one. In English it has sixteen items: twelve single words (um, uh, erm, uhm, hmm, mm, like, so, basically, literally, actually, right) and four phrases (you know, I mean, sort of, kind of). It produces the filler count and the timestamps that the coaching model reasons about.
The wider list, twenty-four patterns, produces the percentage displayed on your report, the one used in the clarity formula and in the community benchmark. Against the English narrow list it adds er, ah, well, okay, yeah, honestly, obviously, anyway, I think, I guess and to be honest. It also misses three sounds the narrow list catches: erm, uhm and mm. So it is wider, but not a strict superset.
The wider list is English, whatever language you spoke. On a recording in another language it will find almost nothing, and the percentage, with the clarity formula built on it, will look better than the delivery was.
The wider list overcounts Several items on the wider list are perfectly good words. "Honestly" can be a real qualifier; "right" can be a sincere question to the room. The wider list cannot tell the difference, so your displayed filler percentage is an upper bound on the problem. The narrower figure is the conservative one.
Four phrases (I think, I guess, kind of, sort of) appear on both the wider filler list and the hedging list. In a live session, where both rates are charged, the clarity score penalises them twice: six points per percentage point as filler, three more as hedging. On a recorded practice take the hedging rate is not charged, so they cost the six points only. Hedging rate has the detail.
What the research says about um
The intuition that fillers are simply damage is not well supported. Disfluencies carry information: an uh before a word predicts that the speaker is about to say something they find hard to retrieve, and listeners use that signal: they recognise the following word faster than they would without it (Fox Tree, 2001).1 Listeners also read a fast, fluent answer as more likely to be correct than a hesitant one, which means disfluency carries real information about confidence (Brennan and Williams, 1995).2
What that research does not license is indifference. The findings concern occasional, meaningful disfluency in conversation. A filler every few words is a different phenomenon, and it costs an audience real attention.
What a filler costs
- 6 pts clarity cost. Per percentage point of filler words.
- 3 pts hedging cost. Per percentage point of hedging language.
- ~16.7% clarity hits zero. About one word in six, with no hedging at all.
How filler and hedging enter the clarity score.
A worked example: a two-minute answer of 280 words with seven fillers has a rate of 2.5%. The formula gives 100 minus 15, or 85. Double the fillers and it gives 70. That formula is one of four inputs to Clarity, alongside the coaching model's own judgement, your pace and your hesitation pauses, as the clarity axis explains.
On the report itself the percentage is put into words: up to 3% reads as Minimal, up to 6% as Occasional, and anything above as Frequent.
What the count misses
It does not know whether a filler was a filler. So opens a great many legitimate sentences, and like is often a verb.
It does not account for how fast you were speaking. A percentage of words is the right normalisation for most purposes, but two speakers with identical rates can sound very different if one clusters their fillers and the other spreads them.
And it does not measure the pause that a filler replaces. The most common reason people produce um is to hold the floor while thinking. Removing the sound without replacing it with a deliberate silence usually makes delivery worse.
Working on it
Use the timeline before the percentage. Find the clusters, then look at what you were trying to say at that moment; a cluster usually marks a transition you had not planned. Rehearse that transition until you know its first words, and let a short silence stand where the um was.
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