AI is the best sparring partner most IELTS candidates will ever have access to, and it is deaf. It will run a Part 3 discussion at two in the morning, challenge every general claim you make, and never get bored of the topic.
It cannot hear that your intonation flattens when you are nervous, or that a consonant cluster is costing you intelligibility. Since pronunciation is one of the four Speaking criteria, that is a quarter of your mark it has no access to.
Knowing exactly where the line sits is what separates candidates who get real value out of AI speaking practice from those who get a false sense of readiness.
What AI genuinely does well
1. Pressure, on demand
Part 3 is where Band 6.5 becomes Band 7.5, and the deciding skill is extending and defending a position rather than deploying impressive vocabulary.
A model instructed to challenge every general claim with a counter-example, press you when you hedge, and never encourage you produces something genuinely close to the real experience. You can run it six times in an evening on six different themes.
The instruction that matters most is the negative one: no praise, no correction, no vocabulary help during the test. Without it you get a friendly conversation, which is the opposite of useful — the absence of reassurance is exactly what produces test-day nerves, and rehearsing without it rehearses the wrong thing.
2. Counting what you cannot count
Record yourself answering a cue card, transcribe it with any free speech-to-text tool, and hand the transcript to a model.
It will give you total words, filled pauses, self-corrections, restarts, your ten most repeated content words, and — the most useful line in the whole output — the grammatical structures a Band 7+ speaker would have used here that you never used once.
| What a transcript can show | What it cannot |
|---|---|
| Filled pauses, restarts, self-corrections | Pronunciation of individual sounds |
| Vocabulary range and repetition | Word and sentence stress |
| Which grammatical structures you used | Intonation and its effect on meaning |
| Whether you answered the question asked | Whether a listener actually understood you |
| How far you developed each answer | Anything about your accent's intelligibility |
The right-hand column is not a minor gap. It is a whole criterion, and it is the one candidates most often assume is fine.
The numbers worth tracking
Language models are unreliable counters — ask twice and the totals shift slightly — so treat these as approximations. Their value is in the trend, not the absolute figure. Record the same cue card today and again in four weeks, run both transcripts through the same analysis, and compare:
- Filled pauses per hundred words. The single most visible fluency marker.
- Restarts and self-corrections. A few are natural and even good; a cluster of them signals you are planning mid-sentence.
- Longest unbroken stretch. Improves faster than almost anything else with practice.
- Distinct content words. A crude but honest proxy for lexical range.
- Structures never used. The gap list, and the best source of drills.
There is no target words-per-minute in the descriptors. Natural IELTS speech usually lands between roughly 120 and 160, but a slower speaker who develops ideas fully scores above a fast one who does not.
The failure mode to watch for
Every consumer assistant is tuned to be encouraging, and encouragement destroys this exercise. The tell is a "great answer!" between questions, or an explanation of what you should have said. The moment either appears, the model has left the examiner role and you are having a language lesson, not a mock test.
Restating the rule mid-conversation rarely fixes it. Put the role instructions in your tool's custom instructions or system prompt before the conversation starts, and begin again.
There is a second, subtler failure in text mode: the model writes the whole test as a script, stubbing your answers as "[User speaks]" and running through to the end without ever waiting for you. Voice mode avoids that entirely, because turn-taking is enforced by the medium.
We document both failures, with the real captured output, on our mock examiner prompt pages.
A routine that works
- Pick a card from our cue card library and answer it in voice mode with the full examiner prompt, recording as you go.
- Transcribe the recording and run the transcript analysis prompt. Save the numbers in a note.
- Run a Part 3 sparring session on the same theme, exactly as the real test does.
- For pronunciation and a band you can plan around, use a tool that processes your actual audio — our Speaking mock test does this on your recording rather than a transcript.
- Repeat the same card in four weeks and compare the two sets of numbers.
If you are preparing entirely on your own without AI in the loop, our guide to practising IELTS Speaking alone covers the offline version of the same routine.