GLOSSARY · A VALIDITY PROBLEM
Why English speech tools misscore Spanish speakers.
The short answer
Most speech tools flag fillers with an English word list. Applied to Spanish, that list charges ordinary words — “este” is a demonstrative (“este año”), “pues” a connective, “entonces” a signpost. So a fluent Spanish speaker is penalized for speaking Spanish normally, and the score reports the opposite of the truth.
The mechanism
An English idea, applied to a language it doesn’t fit.
In English, most fillers are convenient: “um,” “uh,” and “er” mean nothing else, so you can count every occurrence and be right. The word is the filler.
Spanish does not work that way. Its most common “fillers” are also its most common ordinary words. “Este” is the demonstrative in “este año.” “Pues” is a connective. “Entonces” is a signpost — the exact organizing move that good structure rewards. Match them against a flat list and you charge a demonstrative, a connective, and a signpost as if they were “um.”
So the result is not a small error. It is a metric that points the wrong way: the more fluently someone speaks Spanish, the more “fillers” a naive detector finds — because fluent Spanish uses those words more, not less.
The mis-charged words
Every one of these is also an ordinary Spanish word.
| Word | Its ordinary meaning | When (if ever) it’s a filler |
|---|---|---|
| “este” | Demonstrative — “este año” (this year), “este problema.” | A stall only when flanked by a gap: “este… año.” |
| “pues” | Causal connective — “pues no lo sabía” (well, I didn’t know). | A stall only when it stands alone in a hesitation. |
| “o sea” | Reformulation marker — “o sea, es gratis” (that is, it’s free). | A stall when repeated or gapped: “es, o sea, o sea…” |
| “entonces” | Discourse connective — signposting: “Entonces, la solución…” | Not a filler at all — it belongs on the positive, structure side. |
| “nada” | Pronoun — “no dije nada” (I said nothing). | Filler use is a minority; too risky to charge, so it is dropped. |
| “tipo” | Noun — “un tipo de argumento” (a kind of argument). | Regional, young-speaker filler; dropped rather than over-charge it. |
It compounds
The transcriber can be nudged into producing the very words it then charges.
There is a second, subtler layer. Speech tools often prime the transcriber with an example of “what filler-heavy speech sounds like” to help it hear disfluencies. If that example is a string of ambiguous Spanish words, it biases the transcriber toward emitting exactly those words.
Then the same list charges each one. A clean win in English — better filler detection — becomes a compounding penalty in Spanish: the model is nudged into writing down “este,” “pues,” “bueno,” and “nada,” and the speaker is billed for every one. Two mistakes stacked, both pointed at the same speakers.
What a correct detector does
Three tiers, and one lexicon for both languages.
- 01
Count the unmistakable ones directly
Pure hesitation sounds — “eh,” “um,” “uh” — mean nothing else, so they are always counted. No timing needed. This is the part that already works in English, kept unchanged.
- 02
Judge the ambiguous ones by the pause around them
“Este,” “pues,” and “o sea” are counted only when word timing shows a hesitation gap of about a quarter-second on one side, or an immediate repeat (“este… este”). With no evidence of a stall, the honest reading of “este” is the demonstrative. Under-counting a real filler costs a little credit; over-counting tells a fluent speaker their Spanish is a defect.
- 03
Drop what can’t be judged, and credit real structure
Words whose false-positive rate is too high even with timing — “así,” “bueno,” “nada,” “tipo” — are dropped from scoring rather than guessed at. And “entonces” moves to the structure side, because signposting is something to reward, not penalize.
- 04
Match both languages at once
Every session is checked against the English and Spanish lists together. A code-switcher who signposts in English inside a Spanish talk was otherwise scored against half a lexicon — a stable, invisible penalty on exactly the bilingual speakers the tool is for.
This is how SpeakUp Coach scores Spanish. The three tiers above are the shipped behavior, not a wishlist — “eh” always counts, “este / pues / o sea” count only with a real stall, false-positive-heavy words are dropped, “entonces” is treated as structure, and English and Spanish are matched together. Free, in the browser, and you can try it before making an account.
FAQ
Common questions
Why do English speech tools misscore Spanish speakers?
Because they detect fillers with an English word list, then apply it to Spanish. That list charges ordinary Spanish words as disfluencies — “este” is a demonstrative (“este año”), “pues” a connective, “entonces” a signpost. A fluent Spanish speaker gets penalized for speaking Spanish normally, so the filler score can report the opposite of the truth.
Which Spanish words get wrongly counted as fillers?
The usual over-charges are “este,” “pues,” “o sea,” “así,” “bueno,” “nada,” “tipo,” and “entonces.” Every one of them is also an ordinary word: a demonstrative, a connective, an adverb, an adjective, a pronoun, a noun, or a signpost. Counting them the way you count “um” treats normal Spanish as a defect.
Isn’t a filler still a filler in any language?
Some are — “eh” and “um” are pure hesitation with no other meaning, and can be counted directly. The problem is the ambiguous words. In Spanish, whether “este” is a stall or a demonstrative is decided by the pause around it, not by the word. A tool that ignores that context is guessing, and it guesses against the speaker.
How should a filler detector handle Spanish correctly?
With three tiers. Count pure hesitation sounds (“eh,” “um”) unconditionally. Count ambiguous words (“este,” “pues,” “o sea”) only when word timing shows a hesitation gap of about a quarter-second or an immediate repeat. Drop words whose false-positive rate is too high to score at all, and move “entonces” to the structure side. It should also match English and Spanish fillers together, so code-switchers are judged on a full lexicon.
Why is this bias hard to notice?
Because it is stable, not random. A reliability check looks for inconsistency, and this is perfectly consistent — it penalizes the same speakers the same way every time. That looks like a working metric, not a broken one. The people it under-scores are exactly the bilingual speakers a Spanish-capable tool exists to serve, which is the worst place to be quietly wrong.
How does SpeakUp Coach handle it?
SpeakUp Coach uses the three-tier approach above: “eh” is always counted, “este / pues / o sea” are counted only when the timing around them shows a real stall, false-positive-heavy words are dropped, and “entonces” is treated as signposting rather than a filler. It also matches English and Spanish fillers together for code-switchers. All free, in the browser.
Get filler feedback that respects your Spanish.
Record a 30-second speech in Spanish or English and see a filler count that doesn’t punish you for a demonstrative. Free — try it without an account, then a free account (email and password) to keep practicing.
Try SpeakUp Coach — Free