How the Accent Test works
The methodology behind accent-test.com: what one 30-second recording can — and can't — tell you about your English.
By the Accent-Test.com team · Updated 2026-10-24
What the test measures
The test analyses one recording of you reading an English passage aloud (about 20–30 seconds of speech). It reports two separate things.
Pronunciation: how clearly each sound comes through, compared with a reference model of US English, scored from 0 to 100.
Accent similarity: which reference accent groups your voice sounds closest to, with probabilities for the top three.
The two are independent. A strong accent can be perfectly clear, and a native accent can still have unclear words.
From recording to report
- Quality check: the amount of actual speech and the background noise are measured first. With under 8 seconds of speech no accent result is shown; with under 15 seconds, sharing is turned off.
- Speech recognition: a Whisper-family model running on our own servers (with a cloud fallback) transcribes what you said, and the transcript is lined up with the passage word by word.
- Pronunciation scoring: every word is broken into its sounds (phonemes) and each sound is scored 0–100 against the expected pronunciation. Word scores and the accuracy, fluency, completeness and rhythm scores are built from those.
- Accent similarity: two open research speech models compare your voice with recordings of speakers from 14 accent groups and return a probability for each group. A finer-grained accent model is used as a cross-check.
- Report: your scores, word-by-word feedback, the accent result and your two priority sounds.
How the four result levels are decided
The accent result is never just “the top group”. It is one of four levels, decided only from the group probabilities, so you can always see how sure the model is.
| Level | Rule | What it means |
|---|---|---|
| Clear match | Top group at 70% or more and both models agree — or top group at 40% or more and at least 20 points ahead of the next. | One group clearly stands out. |
| Leaning | Top group at 40% or more but not far enough ahead — or at 30% or more when the finer-grained model points to the same group. | A favourite, with the runner-up close behind. |
| Mixed | Two closely related groups (Romance and Slavic, East and Southeast Asia, or two native-English groups) together at 60% or more and within 20 points of each other. | Your English sits between two related groups. |
| Hard to place | None of the above, or less than 8 seconds of speech. | No single group stands out; the highest group is shown with its percentage. |
The 14 accent groups
These are broad families, not countries. Native-English groups: North America, England, Scotland, Ireland, Wales, Australia & New Zealand, and South Africa. Other groups: East Asia, Southeast Asia, South Asia, Romance (Spanish, French, Italian, Portuguese, Romanian), Slavic (Russian, Polish, Czech and others), Germanic (German, Dutch, Nordic) and Semitic (Arabic, Hebrew).
When your top two groups are both native-English groups, the report says “Native-like English” instead of picking a region.
How your two priority sounds are chosen
The report looks at three sound areas that most often cause misunderstandings: final consonants, L vs R, and L vs N. For each area, your scores on the relevant sounds in the passage are compared with typical scores from other learners.
An area qualifies only if your recording has enough examples of it and your average is clearly below typical. The two areas that qualify most strongly — weighted by what is common for your first language — become your two sounds. If none qualifies, the report says so instead of suggesting practice.
Rhythm (word and sentence stress) is practised by everyone in the 28-Day Clarity Program.
What the test can't do
- It doesn't detect nationality, ethnicity or where you live, and it can't tell regional accents inside a country apart (for example New York from Texas).
- It works best on read speech in a quiet room. Spontaneous speech, very short recordings or heavy background noise make it less reliable.
- Pronunciation scores compare you with a US-English reference, so a clear British or Australian pronunciation may differ from it on some vowels.
- Automatic scores are estimates and can be wrong. Treat one recording as a snapshot — scores move a few points between takes.
- The accent models are research models that we keep checking against listeners' judgements; the “Does this sound right?” button on every report feeds that check.
Your recording
The recording from the free test is used only to make your report and is deleted right after the analysis. The report itself (scores and transcript) is kept so you can open it again. Practice recordings in the 28-Day Clarity Program are kept for 30 days and can be deleted at any time. The privacy policy has the details.
Version
Analysis pipeline version 2026.10.24. This page is updated whenever the models or thresholds change.