Filler Words Are Not the Same Word in Every Language
By Noam Shemla · Data & Method
Most public-speaking advice about filler words is written assuming English, and most automated filler detection is built the same way: a fixed list of tokens ("um," "uh," "like," "you know") counted against a transcript. This works fine for English and produces confidently wrong numbers for everything else, because filler words are language-specific rather than a universal, translated set of sounds, and some languages don't use a filler word at all where English speakers reflexively insert one.
Um is a word, and words belong to languages
The most useful idea in this area comes from Herbert Clark and Jean Fox Tree. Working from several large collections of spontaneous English speech, they argued that uh and um are not noise that leaks out when planning fails. Speakers use them to announce a delay: uh for a short one, um for a longer one. Speakers plan them, choose between them and place them much as they would any other word. Their conclusion was that uh and um are conventional English words.1
If that is right, it has a direct consequence for measurement. A word belongs to a language. Other languages have their own conventional items doing the same job, and an English list cannot be expected to find them.
Listeners treat them as words too. In experiments with English and Dutch speech, Fox Tree found that hearing uh helped listeners recognise the word that came next, while hearing um neither helped nor hurt, which fits the idea that the two signal delays of different length.2
Filled pauses differ by language
Linguistic research on disfluency documents that different languages have different conventional filled-pause sounds: English commonly uses "um" and "uh," Spanish speakers commonly use "eh" and "este," French speakers use "euh," Hebrew speakers commonly use "אה" (transliterated roughly as "eh" or "em"), and Japanese has its own set including "ano" and "eto." An English-only detector run on Spanish or Hebrew speech will report an artificially low filler rate that has nothing to do with the speaker's fluency, because it is looking for sounds that speaker never had reason to make.
They also change over time, even between close neighbours. Martijn Wieling and colleagues traced um and uh through spoken and written records in English, Dutch, German, Norwegian, Danish and Faroese, and found the same shift in all of them: um gaining ground on uh, with women and more educated speakers leading the change.3 A filler list is a snapshot of a moving target, to be maintained rather than fixed.
And the conventional sound is only part of it. Much of what a listener hears as filler is ordinary words used to hold the floor: like and you know in English, o sea and pues in Spanish, du coup and en fait in French. These are the hardest to count, because each of them is also a perfectly good word in some other sentence.
Hedging phrases are a second, separate category
Beyond filled pauses, coaching systems also track hedging language, phrases that soften a claim ("I think," "sort of," "maybe," "kind of"). These are even more language- and culture-specific than filler sounds, because hedging is tied to politeness norms that vary across cultures. A list built for English business speech, applied to another language, risks penalising ordinary politeness as if it were a fluency problem. That is a methodology error, not a minor translation gap.
How Exprea handles it
Our timed filler detection selects its list by the language the transcription service detects in your recording, not by your first language. A Hebrew speaker presenting in English is measured against English, because that is the room they are speaking to. There are lists for six languages: English, Spanish, French, German, Portuguese and Hebrew. The Spanish list includes eh, este, pues, bueno and o sea; the French includes euh, ben, du coup and en fait; the German includes äh, ähm, halt and quasi. English is the reference list, and the other five are marked in our own code as provisional, waiting for a native speaker's review. Any other language falls back to English.
One list cannot serve every member
A club whose members speak several languages cannot ship one English word list and call the resulting scores comparable. The sound response is a list per language, built from that language's real disfluency patterns rather than translated from English, and, where a language is not yet well covered, saying so.
What "supported" should mean
A language being "supported" by a speech-analysis pipeline should mean someone built or validated a disfluency list for that language's real patterns. It should not mean merely that the system technically produces a transcript and a number for any audio you feed it. Those are different claims, and conflating them is how a tool ends up confidently wrong in a language nobody checked. By that standard, English is supported here, five more languages are partly supported, and the rest are not.
If you speak in more than one language
Compare yourself only with yourself in the same language. Talks in two languages were counted against different lists, and the report's percentage may barely have counted the non-English one. Listen back instead, and note which words you lean on in each language; they are usually different, and so is the fix.
Built for the case in front of it
This is the same discipline as the Mehrabian piece and the "what we measure" piece in this series, applied to a specific engineering decision: a metric's apparent objectivity is only as good as whether it was built for the case it's being applied to.
References
- Clark, H. H., & Fox Tree, J. E. (2002). Using uh and um in spontaneous speaking. Cognition, 84(1), 73–111. https://doi.org/10.1016/S0010-0277(02)00017-3
- 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
- Wieling, M., Grieve, J., Bouma, G., Fruehwald, J., Coleman, J., & Liberman, M. (2016). Variation and Change in the Use of Hesitation Markers in Germanic Languages. Language Dynamics and Change, 6(2), 199–234. https://doi.org/10.1163/22105832-00602001