Hey Sprout! When a tool tells you "B1, about NCLC 5 to 6", you deserve to know how it got there and how much to trust it. Here is exactly what our free TEF speaking check does with your answer, and what it can't do.
Step by step
1. You answer one question. You get one TEF-style prompt (A2 to B1 level) and speak for 45 to 60 seconds. The recorder accepts up to 75 seconds. If you have no microphone, you can type instead.
2. Silence is trimmed. We cut the silence at the start and end. If there are fewer than 5 seconds of actual speech, we stop and ask you to try again rather than score silence. A failed take doesn't use up your answer or your retry.
3. Speech becomes text. The audio goes to a speech-recognition service set to Canadian French. We get a transcript and the length of your speech, which gives your pace in words per minute.
4. An AI examiner reads the transcript. A large language model gets your transcript, your pace and the question. It is told to act as an experienced TEF oral examiner and a kind teacher, and never to invent things you didn't say. It returns a CEFR estimate (A1 to C1), short notes on grammar, vocabulary and fluency, real corrections, an improved sample answer one level up, and one tip for your retry.
5. CEFR becomes an NCLC range. We map the CEFR estimate to a range, for example B1 to NCLC 5 to 6 and B2 to NCLC 7 to 8. This is a practice mapping we chose. It is not an official conversion, which is why we always show a range and the label "practice estimate, not official".
6. Nothing is stored. The recording isn't kept. The result lives in a signed token in your browser until you choose to save it to an account or email.
What it does well
- Fast, specific feedback. You see in seconds whether your answer is held back by grammar, vocabulary or development, plus a model answer to learn from.
- Pace you can't measure yourself. Many learners speak too little rather than too fast. The pace guide we give the model is simple: natural French is about 120 to 160 words a minute, 90 to 120 is fine for a learner, and under 80 is slow. The advice for slow answers is to say more, not to rush.
- A retry with a goal. One retry, using the tip, so you can feel the difference straight away.
The honest limits
- It reads a transcript, so it doesn't hear your accent. Pronunciation only shows up indirectly: if speech recognition mishears a word, the transcript is wrong, and the feedback can blame a grammar mistake you didn't make. Check the "What we heard" section of the result.
- One question is not the exam. TEF Canada speaking is about 15 minutes with a live examiner: Section A asks you to get information by asking questions, and Section B asks you to convince someone who pushes back. One minute of answer can't show how you handle that.
- The level is a judgement, not a measurement. A language model can be off by a level, especially near a boundary. That's why we show a range and call it an estimate.
- The NCLC mapping is ours. Official NCLC levels come only from your TEF or TCF result, through IRCC's tables.
- AI can be wrong in small ways too. In testing we caught the model listing "corrections" where nothing changed. We now drop those in code and ask for real errors only. We fix issues like this as we find them.
How to use it well
Treat the estimate as a direction, not a grade. Take the check, read the improved answer out loud, use the tip on your retry, and compare. Then practise the skill the feedback points to, and check again on another day.
Try the free TEF speaking check: about a minute, no sign-up. It's a practice estimate, not an official TEF or NCLC result.
Originally published on the FluentSprout blog.

