The short version
- Speakle's scoring engine checks each recording in separate stages: transcription, pronunciation analysis and AI language analysis. No single step decides your result.
- You get scores for pronunciation, fluency, vocabulary, grammar, coherence and staying on topic, plus an overall score, a Confidence Score, a CEFR level estimate and an IELTS-style band estimate.
- These are estimates to guide your practice. They are not official exam results.
- Your recordings are private to you, and you can export your data or delete your account from Settings.
What happens when you record an answer
A standard speaking session is a spoken answer of about a minute to a question or a topic. After you stop recording, it goes through five steps.
1. Your recording is stored privately
The audio is uploaded to private storage inside your account. It is not public. When you play it back in the app, it is served through a link that expires after an hour.
2. Speakle writes down what you said
Speakle's speech engine turns your audio into a transcript and records when each word started and ended. From that timing, Speakle measures how you spoke, as opposed to what you said: your speaking speed in words per minute, the length of your pauses, and how often you used filler sounds. The recogniser also marks words it was unsure about, which matters later.
3. Speakle listens to the sounds
When it is available, Speakle's pronunciation analysis compares the sounds you produced, word by word and sound by sound, and scores how accurately you made them. It uses the transcript from the previous step as its guide. When that analysis cannot run, your pronunciation score is estimated less precisely, from the transcript and from how confident the recogniser was about each word.
4. Speakle's AI coach writes your report
Speakle's AI coach reads the transcript, the timing measurements and any pronunciation results, then writes your report: the scores, what went well, specific mistakes with a correction and a short reason, and one priority to work on next. It follows instructions written to keep the report fair:
- Grammar and vocabulary are judged on the raw transcript, the version closest to what you said, not on a tidied-up version that could hide mistakes or invent them.
- Missing punctuation and capital letters come from transcription, so they are never counted as your errors.
- A mistake that depends on a word the recogniser was unsure about is marked as uncertain, and it must not lower your grammar or vocabulary score. When the evidence is thin, you get the benefit of the doubt.
- Explanations take your first language into account, so a note can explain why a particular slip is common for speakers of your language.
5. Your Confidence Score is calculated
Last, Speakle calculates your Confidence Score with a fixed formula, described below. This step does not use AI, so the same recording always gets the same score.
What each score means
| Score | What it looks at |
|---|---|
| Overall | A summary of the whole answer, from 0 to 100 |
| Pronunciation | How clearly you produced the sounds of English |
| Fluency | How smoothly you spoke: pace, pauses and hesitation |
| Vocabulary | The range and accuracy of the words you chose |
| Grammar | How accurately you built your sentences |
| Coherence | Whether your ideas connected and were easy to follow |
| On topic | Whether you answered the question you were given |
| Clarity | One of four bands, from easy to understand to hard to follow |
| Confidence | How confidently you delivered the answer, explained below |
| CEFR estimate | Your likely level on the six CEFR levels, from A1 to C2 |
| IELTS-style estimate | A band estimate on the 0 to 9 scale used for IELTS speaking |
How the Confidence Score works
The Confidence Score measures delivery, never correctness. Grammar, vocabulary, pronunciation accuracy and your CEFR level are deliberately left out. A beginner who speaks steadily with simple mistakes can score higher than an advanced learner who waits a long time before a perfect sentence, and we think that is right.
It combines several parts of your delivery, mostly taken from the timing of your recording:
- Starting: how quickly you began to speak
- Flow: how long you kept going without long breaks
- Hesitation: how often you paused, used fillers or hedged
- Pace: whether you kept a steady speaking speed
- Staying in English: whether you carried on in English when you got stuck
- Expressiveness: whether your intonation was flat, when that measurement is available
- How it felt: your own rating after you speak, if you give one, which can count for no more than a tenth of the score
The result falls into one of five bands: Finding your voice, Warming up, Steady, Assured and Commanding. When a recording is too short to judge, no Confidence Score is recorded for it.
Some automated measures of fluency and intonation are built around how close someone sounds to a native speaker. We leave those out of the Confidence Score, because a confident speaker with a strong accent should not lose points for the accent.
How CEFR and IELTS-style estimates are made
CEFR level. The Common European Framework of Reference for Languages describes ability on six levels, from A1 to C2. The AI coach places your answer on that scale from the grammar, vocabulary and organisation of what you said. It is told to judge language, not accent: a strong accent must not lower a level that your grammar and vocabulary support. A minute of speech is a small sample, so when an answer sits between two levels it is told to choose the lower one, and a one-minute placement answer is never given C2.
IELTS-style band. Your report also gives a band estimate on the 0 to 9 scale used for IELTS speaking. To keep the two estimates consistent, the band is kept inside the range that published alignments between CEFR levels and IELTS bands give for your estimated CEFR level. At A1 and A2 that alignment is approximate, because IELTS is not designed to tell those levels apart finely.
Limits you should know about
We would rather tell you the limits plainly than let a number look more certain than it is.
- These are estimates, not certificates. Speakle is not an exam board. Its CEFR and IELTS-style estimates are not official results from IELTS, Cambridge or any other test provider, and they cannot replace a real exam.
- One answer is a small sample. Your score can change with the topic, how tired you are, your microphone and background noise. Look at your trend over several sessions rather than at any single result.
- Speech recognition can mishear. Noise, a quiet voice or an accent the recogniser is less used to can put wrong words in the transcript. Uncertain words are flagged so they do not count against you, but errors can still slip through.
- Pronunciation scores compare you with reference speech. They can mark down an accent even when you are easy to understand. We treat pronunciation as a tool for being understood, not for sounding like a particular kind of speaker.
- AI feedback can be wrong. The AI coach can occasionally misjudge a sentence or suggest a correction that was not needed. If something in your report looks wrong, trust your own judgement and tell us at [email protected].
The correction dial
Speakle lets you choose how much correction you see:
- Light: only what stops someone understanding you, one thing at a time.
- Balanced: the one or two things that matter most. This is the usual setting.
- Strict: everything, including phrasing that is correct but not what a fluent speaker would usually say.
The dial changes what is shown to you, not what is analysed. The same analysis runs at every setting, and every mistake is still saved to your progress. After a speaking session, the report shows up to one mistake on Light, up to three on Balanced and up to five on Strict, and suggestions for more natural phrasing appear only on Strict. The full picture is on your Progress page whenever you want it, and you can change the dial at any time. In live conversations, the dial also tells the AI coach how much to correct while you talk.
Feedback in your own language
Explanations and hints can be shown in nine languages: Arabic, Hindi, Japanese, Korean, Malay, Thai, Vietnamese, Turkish and Portuguese. Notes about your first language are based on well-known patterns in how speakers of each language tend to carry sounds and grammar into English, drawn from established pronunciation teaching material. They are a starting point for your feedback, not a measurement of you.
What happens to your recordings
- Private to you. Nobody else can play your recordings. They are kept in private storage inside your account.
- Used only to score your speaking. Your audio and transcript are processed by Speakle's scoring engine and the specialist processors it runs on, each bound by a written agreement to use it only for this. Our Privacy Policy describes each category of processor.
- Not sold, and not used to train other companies' models. We do not sell your data, and we do not use your personal practice data to train third-party AI models.
- Kept so you can hear your progress. Recordings stay in your account so you can compare your first answers with your latest ones.
- Yours to take away or delete. You can export your data or delete your account from Settings. After you ask to delete your account there is a 30-day grace period, in case it was a mistake, and then your personal data is permanently erased. Billing records are kept for longer where tax law requires it.
For the full details, read the Privacy Policy and the Data Policy. Questions about scoring are welcome at [email protected], and questions about privacy at [email protected].
