How our AI grades your writing and speaking
Every written and spoken answer is rated three times by an AI rater against a four-part rubric, with evidence quoted for each level. The median becomes an estimate on the 10–160 scale, shown with a range, plus feedback you can act on.
- Four parts rated 0–6: task, grammar, vocabulary, organisation
- Speech is transcribed; speed and pauses come from the audio timings
- An estimate with a range, never presented as an official score
What happens to an answer after you submit it
1. Speech becomes text (speaking only)
Your recording goes to a speech recogniser (Deepgram), which returns the words and the time each was said. The audio is then discarded; we never store it.
2. Three independent ratings
An AI rater reads the task and your answer and rates four parts from 0 to 6. For each part it must first write one sentence of evidence from your answer, then give the level, so the level follows from the evidence.
3. The median of the three
For each part we take the middle of the three ratings, so one unusual rating cannot move your score. If the three disagree, your score range widens.
4. Task fulfilment caps the rest
Your overall level cannot be more than one and a half levels above your task-fulfilment level. Fluent English that does not answer the question cannot score highly.
5. Fluency, measured in code (speaking only)
Words per minute, pauses of half a second or more and filler words come from the timings, not from the AI. They nudge the speaking estimate up or down, within a small fixed limit.
6. A score with a range
The combined level becomes an estimate on the 10–160 scale. The range is wider when your answer was short or the ratings disagreed.
The four parts, and what each level means
| Part | What the rater looks for |
|---|---|
| Task fulfilment | Does the answer address the prompt, with relevant content, at a sensible length? |
| Grammar | Accuracy and range of structures (in connected speech, for speaking). |
| Vocabulary | Precision, range and how well the words fit. |
| Coherence | Organisation, logical flow and linking. |
| Level | What it looks like |
|---|---|
| 0 | No usable response, or entirely off topic. |
| 1 | Isolated words or memorised fragments; meaning mostly lost. |
| 2 | Very simple sentences; frequent errors that block meaning (about A1). |
| 3 | Simple connected text on familiar topics; errors frequent but meaning usually survives (A2–B1). |
| 4 | Clear, connected text with some complex structures; errors noticeable but rarely block meaning (B1–B2). |
| 5 | Fluent and well organised, varied structures and words; minor errors (B2–C1). |
| 6 | Precise, nuanced, naturally organised; near-native control (C1–C2). |
The rater never sees the 10–160 scale. It only places your answer on these levels; our code does the arithmetic. That keeps it from drifting towards round numbers.
What you see on your report
| On every graded answer | Why it is there |
|---|---|
| The level for each of the four parts, with the evidence | So you can see exactly why, not just a number |
| Your strongest feature | To keep doing it |
| The one thing to work on next | The highest-value fix, in one sentence |
| A note on length | Too short is the most common avoidable loss |
| Speaking: words per minute, long pauses, fillers | The fluency side of your speaking score |
See it on a sample report, or take a full mock test to get your own.
What it does not do, and why the score is an estimate
- It is not Duolingo’s grader. Duolingo’s own engine is not public. Ours is built independently to measure the same skills on the same scale, so the result is an estimate, labelled as one on every report.
- It does not hear your accent. The rater reads a transcript. The recogniser’s confidence is used only as a rough sign of clarity.
- No person grades your answers. That makes it instant and consistent, and it means there is no human judgement in the loop.
- Short answers are less certain. Fewer words means less evidence, so the range is wider.
Questions about our AI grading
How does DET Mock Tests grade writing?
An AI rater reads your answer three times, independently, and rates four things on a 0–6 scale: whether you answered the task, grammar, vocabulary and organisation. It must quote evidence from your answer before giving each level. We take the median of the three ratings for each part, combine them, and turn that into an estimate on the 10–160 scale with a range.
How does DET Mock Tests grade speaking?
Your recording is turned into text by a speech recogniser, and the text is rated on the same four parts as writing. Separately, the word timings give your speaking speed, long pauses and fillers, which can move the estimate up or down a little. The audio itself is never stored.
Which AI grades my answers?
A large language model from Anthropic (Claude) rates the writing and the transcripts, and Deepgram converts speech to text. Both are named on our privacy page. The rubric, the three-rating median and the conversion to a score are our own code.
Why is each answer rated three times?
Because any single AI rating can be off. Three independent ratings with the median taken per part means one unusual rating cannot move your score, and when the three disagree, your score range is widened to show the extra uncertainty.
What are the four things my writing and speaking are rated on?
Task fulfilment (did you answer the question, at a sensible length), grammar (accuracy and range), vocabulary (precision and range) and coherence (organisation and linking). Each gets a level from 0 to 6, from no usable response to near-native control. The grader writes down the evidence for each part before it chooses the level.
Why did a well-written answer get a low score?
Most likely it did not answer the question that was asked. Task fulfilment caps the rest: the overall level cannot be more than one and a half levels above it. Beautiful English about a different topic is not evidence of the skill being tested, so it scores low by design.
Is my pronunciation scored?
Not directly. The rater sees a transcript, not your voice. Speaking speed, long pauses and fillers are measured from the audio timings, and the recogniser’s confidence is used only as a rough sign of how clear the speech was. That is why clear, steady speech matters more than any particular accent.
Does a person ever read my answers here?
Not to grade them. Grading is fully automated, so it is instant and the same for everyone. Your written answers and transcripts are stored with your attempt so you can reread the feedback, and they are deleted if you delete your account; audio is never kept.
How accurate is the AI grading?
It is an estimate, and we show it as one: every score comes with a range, and the range is wider for short answers or when the three ratings disagree. We check the grader against reference answers at fixed levels, and we ask students who sit the real test to tell us their official score so the estimates can be compared and adjusted.
Is this the same way the real Duolingo English Test is graded?
No. Duolingo uses its own grading engine, which it does not publish in detail. Ours is built independently to measure the same skills on the same 10–160 scale, using a published-style rubric. That is why our scores are estimates, not official results, and we say so on every report.
Why is my score shown as a range?
Because a single number would claim more precision than any practice test has. The range comes from the uncertainty of the ratings: how long your answer was, and how much the three ratings and the four parts agreed. A narrow range means the evidence was clear.
What feedback do I get on each answer?
The level for each of the four parts with the evidence behind it, your strongest feature, the single most useful thing to work on next, and a note on length. For speaking you also see your words per minute, long pauses and fillers.
Can I trick the grader with memorised answers?
Not usefully. Memorised material that does not answer the actual prompt is capped on task fulfilment, and the rubric is fixed on the server, so nothing you type can change how it is applied. The best strategy is the honest one: answer the question clearly.
Does the length of my answer matter?
Yes, in two ways. An answer that is too short for the task loses on task fulfilment, and very short answers give the rater less evidence, so the score range is wider. Longer is not automatically better, though: relevance and control matter more than word count.
What happens to my recordings?
They are sent to the speech recogniser to be turned into text, and then discarded. We never store audio. Only the transcript and the feedback are kept with your attempt. That means nobody, including us, can play your answers back later; what remains is the text of what you said and the feedback on it.
How can I help make the grading more accurate?
If you have taken the real Duolingo English Test, tell us your official score and the date (through the feedback pop-up or the contact page). Comparing real scores with our estimates is how the engine gets calibrated. Even one real score with its date helps, because it shows how far our estimate was from the official result.
AI grading: quick reference
| Question | Answer |
|---|---|
| Who grades | An AI rater (Anthropic’s Claude); speech is transcribed by Deepgram. |
| Rubric | Task fulfilment, grammar, vocabulary, coherence, each 0–6. |
| Ratings per answer | Three, independently; the median is used. |
| Off-topic answers | Capped by task fulfilment. |
| Speaking fluency | Words per minute, long pauses and fillers, from the audio timings. |
| Audio | Never stored. |
| Result | An estimate on the 10–160 scale, with a range. |
| Official? | No. Only Duolingo gives official scores. |
Start a practice test
An hour under real conditions, all eight subscores, and feedback on every written and spoken answer.
Take a test