A spoken answer the recognizer did not catch is not a wrong answer. Quizlar asks once more instead of scoring it.
Checked against production code on .
In short
Outcome
An answer the recognizer did not catch is handed back to you, not marked wrong. Your review schedule only moves on answers that were heard.
Cost
One extra sentence, a few seconds, at most once per reason per question.
We guarantee
A card that could not be heard after a repeat is left unscored, with nothing written to your review history. Every prompt quoted on this page is the one the app speaks.
We do not guarantee
That every mishearing is caught. Where it is not is listed on the accuracy and limits page. No speech-recognition accuracy figure is claimed anywhere on this site.
The fear
You close the notes and say the answer out loud. There is nobody in the room to tell you whether you had it. The question you have about any voice grader is the obvious one: what if it just does not hear me?
It matters more than in a typed quiz. Each grade feeds the review schedule. A wrong mark on an answer you knew pulls the card forward for no reason: it comes back tomorrow when it should have come back next week.
The rule
A transcript the grader cannot trust is not graded. It is handed back to you with a request to say it again. The grade only lands on an answer that was heard well enough to judge.
Each reason below fires at most once per question. After that the answer is graded as heard, with one exception: a card the recognizer still cannot catch is left for next time, not scored.
A grader that scores a submitted recording has no moment at which to ask. Quizlar grades a live turn, so it does.
When it asks
Low recognition confidence. The speech recognizer reports how sure it is about the words. Below the floor you hear: "I didn't quite catch that — could you say your answer again?" The floor exists because of a real session: a spoken "kijani" came back as "Okay." at low confidence and was about to be graded as the answer. Only some speech providers report a confidence that means word accuracy; see the limits page for which.
A cut-off answer. If the transcript is three or more words and ends on a dangling word such as "and", "the" or "of", the recognizer probably closed the turn early. You hear: "It sounds like your answer got cut off — could you say your full answer again?" Typed answers skip this check; they were submitted whole.
A near-miss. What was heard sounds like the right answer but is not close enough to accept and not clearly something else. Rather than guess, the grader asks: "I didn't quite catch that — could you say your answer again?" English only: this check compares the sounds of words, which the matcher cannot do in other languages.
Unclear pronunciation. The words were aimed at the right answer but came out mangled, "perris" for Paris. You hear: "Good — now say it once more, nice and clear." This is a prompt to practise, not a doubt about your knowledge; a small slip like "asubui" for "asubuhi" does not trigger it.
When it asks something else
An ambiguous answer gets a question, not a repeat. If the model grader reads your answer as maybe-right, it asks once: "Did you mean X?" or "X or Y?" A bare yes or no settles it. Anything else is graded with the question as context.
A wrong answer that is close may get a guiding question before the grade, a Socratic hint. One follow-up per card, whichever kind fires first.
What to do
Say the whole answer again as a complete phrase, not a fragment. For a list, say every item.
If it still cannot hear you, type it. Every question has a "Type your answer" field under the mic. A typed answer is graded as typed.
If you were asked for the same reason and it still could not catch the words, the card is skipped: "I still couldn't catch your answer, so I'm leaving this card for next time. The correct answer is: …" No grade, no rating, nothing written to your review history. The schedule stays where it was.
What it costs you, and what it does not promise
Cost: one extra sentence, a few seconds, at most once per reason per question.
Not promised: that every mishearing is caught. Some speech providers report no usable confidence, so a mishearing from them goes straight to grading. The full list is on the accuracy and limits page.
Not claimed: a recognition accuracy figure. We do not publish one because we do not have a measurement that would hold across subjects, accents, microphones and 60+ languages.