Skip to main content
filler word tracking runs locally on every transcript segment. it counts exact matches of single-word and multi-word phrases against a per-language list. no llm, no api calls.

​
how miniti counts them

  1. the transcript is tokenized (lowercased, punctuation stripped, apostrophes preserved)
  2. each filler from the active language list is matched as a sequence of tokens - so “you know” matches two adjacent words, not the word “you” or the word “know” on their own
  3. only final transcript segments are counted. deepgram’s interim results don’t contribute, so you won’t see false spikes while a sentence is still forming
  4. counts are per-speaker and per-minute (fillers-per-minute = fillers / session minutes)

​
default filler lists per language

these are the out-of-the-box lists. you can override any of them in settings.

​
english

um, uh, hmm, hm, er, ah, like, basically, literally, actually, honestly, uh huh, you know, i mean, kind of, sort of

​
spanish

eh, este, bueno, o sea, pues, es que, digamos, entonces, a ver

​
french

euh, ben, genre, en fait, du coup, voilà, quoi, bah, bon

​
german

äh, ähm, halt, also, sozusagen, quasi, irgendwie, na ja, genau

​
portuguese

é, né, tipo, assim, então, bom, quer dizer, enfim

​
italian

ehm, cioè, tipo, allora, praticamente, insomma, diciamo, boh

​
dutch

eh, uhm, eigenlijk, zeg maar, weet je, dus, nou, gewoon

​
swedish

eh, öh, liksom, typ, alltså, asså, va, ju, ba

​
greek

ε, εε, δηλαδή, κοίτα, λοιπόν, ας πούμε, τέλος πάντων

​
polish

ee, no, w sumie, jakby, znaczy, generalnie, w zasadzie, tak naprawdę

​
russian

эм, ну, вот, типа, короче, как бы, в общем, значит, так сказать

​
customizing the list

default lists are a starting point - what counts as a filler depends on your context. settings → training → filler words, then pick a language. two reasons to customize:
  1. add your personal tells - “kind of”, “right?”, “so”, “basically” are common repeats that aren’t always in the default list. if your saved meetings keep flagging you on a word you never noticed, add it.
  2. remove false positives - “like” is a filler most of the time but a valid verb (“i like that”). “actually” is sometimes meaningful contrast. if the metric over-penalises you, remove words that your context uses meaningfully.

​
what to add first

listen back to one saved meeting in full. every time you skip past yourself saying the same thing twice in 30 seconds - that’s a candidate. common additions people don’t realise:
  • bridging tics: “so”, “right”, “okay”, “yeah”
  • softeners: “i guess”, “i suppose”, “maybe”, “probably”
  • hedges: “a little bit”, “somewhat”, “fairly”
  • verbal stalls: “one second”, “let me think”, “how do i put this”

​
reset to defaults

if you go too far, reset in settings restores the language’s default list.

​
how to actually reduce fillers

fillers usually spike when you’re buying time to think or starting a sentence you haven’t finished composing. three practical interventions:

​
1. the deliberate pause

replace “um” with silence. a 1-2 second pause feels long to you but short to the listener - and it sounds like confidence, not hesitation. the training metric drops immediately because silence isn’t a filler.

​
2. start sentences you’ve already formed

most fillers happen in the first 3 words of a sentence because you started before the sentence was ready. take half a breath, compose the sentence in your head, then open your mouth.

​
3. end sentences on the word, not the trailing filler

“…so that’s what i was thinking, you know?” has two fillers. drop them: “…so that’s what i was thinking.” the metric will thank you; the listener will thank you more.

​
measuring improvement

  • per meeting: check fillers/minute in the training pane for each saved meeting
  • per type: fillers that show up on discovery calls might not show up on internal standups, and vice versa. compare like-for-like
  • per week: aim for a trend, not a target. one noisy meeting in a good week is normal
see tracking progress for the full workflow.

​
caveats

  • fillers require exact token matches, so “ummmm” (not in the list) won’t count the same as “um”. add variant spellings if your speech pattern includes them
  • partial matches don’t fire - “basic” won’t count as “basically”
  • multi-word phrases (like “you know”) must appear as adjacent tokens with no punctuation between them
  • interim transcript text is intentionally ignored to avoid flicker - fillers are counted when the segment is finalised

​
see also