CELPIP Speaking Checker (AI Practice Feedback)
Record or upload an answer to any of the 8 tasks. Get an estimated level range, timestamped evidence from your own words and two rewrites.
Task 1 · Giving Advice
Making friends in a new city
Your friend Maya has just moved to a new city for work. She is unsure how to meet people outside her workplace. Give her advice about practical ways to make friends and feel more connected to the community.
Quick tips for this task
- Speak to the person directly and acknowledge their situation.
- Offer two or three concrete suggestions, each with a reason.
- Close warmly — 90 seconds is longer than it feels.
Ready
- Prep
- 0:30
- Speak
- 1:30
Your browser will ask for the microphone. Recordings are scored, then discarded — never stored.
Yes — as a practice estimate. Choose the task and prompt you answered, then record or upload your raw answer. The checker transcribes it, estimates a level range on the four CELPIP criteria with quoted, timestamped evidence, and suggests what to fix first. It is not an official score and can’t replace trained CELPIP raters.
What score would my answer get, and why?
You get a range, never a single number, and every judgement points to your own words: a quote, the criterion it affects and the time in the recording. Click a quote to hear that moment. The priorities list then turns the evidence into one or two changes to make before your next attempt.
- Evidence → criterion → change. “which is a great way to meet neighbours” → Content & Coherence → add a specific example.
- Timed transcript. Words light up as the recording plays; fillers and long pauses are marked with a legend, not colour alone.
- Hard rules applied first. Very short answers (under 40% of the time or 50 words) are capped, and the report says so.
- One task, not your whole test. A single answer can’t predict your full speaking result.
Which criteria does the checker use, and is my problem pronunciation, fluency or content?
The four published CELPIP speaking criteria: Content & Coherence, Vocabulary, Listenability and Task Fulfillment, each with its own range and evidence. Pronunciation, rhythm, pausing and grammar sit under Listenability; ideas and order under Content; answering the task for the right listener under Task Fulfillment.
| If the report shows… | It usually points to | Try this next |
|---|---|---|
| Many long pauses and restarts | Fluency and organisation (Listenability, Coherence) | Plan key words only; speak in meaning groups |
| Clear delivery but few reasons or examples | Content & Coherence | Give every point a reason and an example |
| Good ideas but the wrong listener or tone | Task Fulfillment | Re-read who you are talking to and why |
| General, repeated words | Vocabulary | Swap two general words for precise ones |
| Grammar slips in time or reference | Listenability (grammar) | Fix one repeated pattern, then test on a new prompt |
For the full symptom-to-drill table, see Is your problem pronunciation, fluency, grammar or content?
Can AI rewrite my answer at a higher level that still sounds like me?
Yes. You get two versions of your own answer: a minimal edit that keeps your words and fixes only what matters, and a target-level version with the same ideas, more specific detail and stronger linking. Each change has a reason. Then practise saying it aloud under the timer — not memorising it.
An invented short example (not anyone’s real answer, and not a full test response):
I think bus good. It cheap and easy.
I think taking the bus is a good choice because it’s cheaper and you don’t need to find parking.
If I were you, I’d try the bus for a week before buying a car. It would keep your costs down, and you could use the trip to read or relax. Just check the evening timetable first, especially if you sometimes work late.
The higher version gains specific advice, reasons and a practical limit — not rare words. Re-record it on a new prompt to check it transfers.
What can’t AI judge reliably?
- Your official level. Official results come from trained raters and CELPIP’s own scoring process across all tasks.
- Pronunciation from text alone. If the model only receives the transcript, the report says pronunciation wasn’t assessed.
- Poor recordings. Noise, clipping or a distant mic lower confidence; re-record before acting on the result.
- Accent. A clear accent is not an error. The checker looks at how easy you are to follow.
We publish our method and calibration plan on AI scoring accuracy.
Is it free, and is my recording stored?
You can record, listen back and read every sample answer for free. AI reports use report credits; the current free allowance is listed in numbers on our pricing page, and plans are one-time purchases that don’t renew.
Audio is processed for scoring and then discarded — we keep the transcript and the report, not the recording. Reports are private to your account or browser and you can delete them.
Questions people ask
Which files can I upload?
Can I check an answer to my own question?
How accurate is the estimate?
Next stations
Unofficial practice and estimates — not affiliated with Paragon or IRCC, and not legal advice. Read the disclaimer