Method and accuracy
How WitAccent works, and how accurate it is
Written and reviewed by Fatih Can YILDIRIM, founder of WitSpeak ·
WitAccent listens to about 30 seconds of read-aloud English and compares it with the eight accent groups it was trained on. On a research test set it never trained on, it picked the right group 76.2% of the time. On real browser recordings it is much less sure, so most of the time it holds back instead of guessing. Below are the dated numbers, what they cover, and what they don't.
What happens when you take the test
- 1
You read one short English passage aloud in your browser. The recording goes only to WitSpeak's server, through witspeak.ai.
- 2
A speech check runs first: recordings that are too short, silent or too noisy are turned away before any guess.
- 3
A speech encoder turns the audio into an acoustic representation, and a small classifier compares it with eight groups: Chinese (Mandarin), South Asian, Filipino, Turkish, Spanish, Arabic, Russian, and native-like English.
- 4
A guardrail decides what you see. A non-native group is named only when its confidence is at least 0.55, and "native-like English" is never shown as a verdict. Otherwise WitAccent says it isn't sure, and the result screen may show a hint based on your connection's country or browser language, always labelled as a location-based estimate.
- 5
You can tell WitAccent your first language. Agreement is measured only with answers given before the guess is revealed, because seeing a guess can change the answer.
Measured accuracy (research test set, model v3, August 29, 2026)
These numbers come from a frozen test set of 5,626 recordings that were never used for training.
- Top-1 accuracy
- 76.2%
- How often the right group came first
- Macro-F1
- 0.30
- Average across groups; low because the small groups score poorly
- Calibration error (ECE)
- 0.092
- Gap between stated confidence and real accuracy; lower is better
- Held back at the live threshold
- 18.3%
- Accuracy when it does name a group: 83.7%
Most of the 76.2% comes from Mandarin-Chinese and native-English clips (4,807 of 5,626). The model is strong on those two groups; on the others it is weak or not yet measurable:
| Group | Clips in the test set | F1 |
|---|---|---|
| Chinese | 3,352 | 0.92 |
| Native-like English | 1,455 | 0.74 |
| South Asian (Hindi, Urdu…) | 232 | 0.05 |
| Filipino * | 50 | 0.11 |
| Spanish * | 21 | not measurable |
| Arabic * | 8 | not measurable |
| Turkish * | 0 | not measurable |
| Russian * | 0 | not measurable |
* Training data for Filipino, Turkish, Spanish, Arabic and Russian is mostly synthetic speech, so even a high score for these groups could reflect the voice source rather than the accent.
How it does on real-world recordings
- Public crowd-sourced accent clips (Common Voice, mp3)
- 48.7%
- Top-1 accuracy on 1,687 clips
- Recordings made in the WitSpeak app
- 0 / 35
- Only 3 speakers, so a small sample, but a clear warning that phone audio differs from the training data
- Live web tests (August 31, 2026 – October 25, 2026)
- 1,655
- Named a group in 323 (19.5%); held back in 1,332 (80.5%)
- Agreement with a first language given before the reveal
- 8 / 68
- Tests where someone answered before seeing the guess and WitAccent named a group
Why does it hold back so often on the live site? Browser recording conditions differ from the research data, and of the 481 first-language answers, 229 named a language outside the seven groups and 56 said English. Holding back when it isn't sure is the designed behaviour, not an error.
See all the aggregate numbers →What WitAccent cannot tell you
- It does not verify your first language, nationality or identity.
- It does not score your pronunciation, fluency or English level.
- It only knows seven non-native groups; any other first language can only produce "not sure" or a wrong group.
- Multilingual backgrounds, years abroad, the microphone and room noise all change the result.
- No comparison with human listeners has been run yet; that cell is unmeasured.
Your recording and data
- Taking the test doesn't require contributing data: we process the recording to return your result.
- We keep a recording to improve the model only if you separately choose to contribute (adults only, not for school use). Contributed recordings are deleted after 30 days, and you can delete a contributed result yourself.
- A keyed hash of your IP address is used only for abuse limits. A coarse country code may be kept for aggregate sample balancing, never as an accent label or a model input.
Who wrote this
Written and reviewed by Fatih Can YILDIRIM, founder of WitSpeak. Accuracy figures come from the v3 model scorecard of August 29, 2026; live figures from the WitAccent data snapshot refreshed on October 25, 2026. We update this page when the model or the threshold changes.