How your English speaking score is calculated

The standards we follow

Our scoring follows how spoken English is assessed by official exams and by the best-known automated speaking tests. We didn’t invent new skills; we took the ones examiners already use.

Pronunciation: how clearly you say each sound

Pronunciation is measured from your voice, not from the transcript. Each answer you record is checked by Microsoft’s pronunciation assessment, which scores every sound (phoneme) against how a native speaker says it, in the American or British accent you chose.

• Sound accuracy (70%): how closely your sounds match, averaged over the call. Longer answers count more than one-word ones.
• Intonation (30%): stress, rhythm and the rise and fall of your voice, where it can be measured (American English).

One adjustment: the raw measurement sits high for anyone who is understandable. In our first calls, learners from A1 to B1 all got between 82 and 87, so the number barely told them apart. We spread it over the full scale (a raw 50 becomes 0, a raw 97 becomes 100), so a real difference in clarity shows up as a real difference in your score.

Your report also shows every sentence and word, coloured by how clear it was, and lets you practise the hardest words. Those detailed percentages are the raw measurement.

Fluency: pace, pauses and speaking at length

Fluency is measured from the timing of your words. Research on second-language speaking finds that two timing measures track a speaker’s level best: speech rate and mean length of run, the average number of words you say between pauses. We use both, plus how much you say per answer.

Answer length matters because very short replies are easy to say quickly. Without it, someone who only says “yes” and “I like it” could look more fluent than someone explaining a real story.

Grammar: accuracy and range

Like IELTS, we score grammar on two things: how correct it is and how much grammar you use. Simple, correct sentences are good, but they don’t score like complex ones.

• Accuracy (50%): an AI reviews the transcript and marks every real error, “minor” or “major” (major only if it changes or hides the meaning). Each error costs 3 points per 100 words you said; a major error counts double. On short calls the score is pulled toward what your level usually scores, so 20 error-free words can’t give a perfect score.
• Range (50%): the AI rates the grammar you used against CEFR descriptions, from A1 (memorised phrases, simple present) through B1 (connected sentences, present perfect, comparisons) to C1–C2 (a wide range of complex structures). The rated level is worth these points:

Example: 80 words with 3 minor errors and 1 major error (5 weighted errors, 6.25 per 100 words) gives about 80 for accuracy. With B1 range (60), grammar is 70.

Vocabulary: range and less common words

Vocabulary is half rated, half counted.

• Range (50%): the AI rates your vocabulary against CEFR descriptions, from everyday words for routine matters (A2) to varied, precise words with little searching (B2) and natural idiom (C2). Same points as for grammar.
• Beyond the basics (50%): the share of your words that are not among the few hundred most common English words (like “go”, “good”, “thing”). 10% is 0 and 45% is 100. In our first calls, A1–A2 learners used about 20% and B1 learners about 30%.

Example: 80 words, 20 of them beyond the basics (25%), with B1 range: vocabulary is 51.

We tried measuring variety instead (how many different words you use), but on calls of about 50 words almost everyone scored the same, so it didn’t help.

From four skills to one score and a CEFR level

Your overall score is the average of the four skills, each worth 25%, as in IELTS. If a skill couldn’t be measured (for example, no audio was recorded), the others share its weight.

Your level is then read from the overall score with fixed bands, the way Pearson turns its test scores into CEFR levels. So you’ll never see a high score next to a low level, or the other way round.

We set these bands on our first real calls so that the level from your score matched the AI’s overall judgement of each call. They agreed on most calls and were never more than one level apart, which is about how often two human examiners disagree. The A1–B1 bands rest on real calls; the higher bands follow the same spacing and will be adjusted as more people use the app.

A full example, and try it yourself

What the score can and can’t tell you

• It’s for practice, not a certificate. Official exams like IELTS use trained examiners and set tasks; use them when you need a recognised result.
• Short calls are less reliable. We need at least 20 words to give a score, and a few minutes of real answers give a much better one. The home screen averages your last three calls for the same reason.
• Your accent is not a mistake. Pronunciation measures how clear each sound is, not whether you sound like a native speaker. Clear speech with an accent scores well.
• The AI can be wrong. It finds errors and rates range, but the timing and sound measurements don’t depend on it, and each error in your report is shown with its correction, so you can check.
• We’ll keep improving it. The calibration uses our first calls. As more people practise, we’ll check it again, especially for advanced speakers, and update this page.

How to raise each score

• Pronunciation: practise the words your report marks in red, and the sounds learners find hardest: the TH sound, word stress, linking sounds.
• Fluency: answer in full sentences and add a reason or an example; don’t stop after a few words. More ideas in how to speak English fluently.
• Grammar: fix the errors your report lists, then try one new structure per call, such as “I’ve never…” or “If I had…”.
• Vocabulary: use the new words your report suggests in your next call, and replace basic words with precise ones (“delicious” for “very good”). Start with daily use words.