AI IELTS Speaking Examiner: What It Can and Cannot Do

What it measures well

The four IELTS Speaking criteria each contain features that can be measured from a recording, and this is where an automated examiner is reliable.

• Fluency markers. Pause frequency and length, speech rate, repetitions, false starts and fillers ("um", "like"). A human examiner forms an impression of these; software counts them, and it counts them the same way every time.
• Pronunciation patterns. Consistent substitutions (/v/ for /w/, /b/ for /p/), dropped final consonants, flat intonation and misplaced word stress. The model can tell you that you said "development" with the stress on the first syllable nine times, which no human has the patience to log.
• Lexical range. How many distinct words and collocations you used, how often you repeated a word, whether less common items appeared and whether they fitted.
• Grammatical range and accuracy. The mix of simple and complex sentences and the proportion that contained an error, which is exactly what the descriptor asks the examiner to judge.
• Relevance. Whether the answer addressed the question asked, which is the Part 3 failure most candidates do not notice in themselves.

Because these features are counted rather than felt, the AI is consistent in a way human examiners are not. Two candidates with the same pause pattern get the same fluency note, and the same candidate on Monday and Friday gets a comparison that means something.

What it cannot judge

Three things stay human. Interaction: whether you asked for clarification at the right moment, whether you picked up on the examiner's cue to extend, whether you sounded like someone in a conversation. Nuance: irony, humour, understatement, the register shift between Part 1 chat and Part 3 argument. Communicative effect: whether a native listener would find the answer persuasive or merely correct. An automated examiner can report that an answer was fluent, accurate and on topic; it cannot report that it was dull, and in Part 3 that distinction can be half a band. The IELTS Speaking band descriptors explained shows where each of these human judgements enters the official criteria.

How Langujet scores speaking

Langujet's speaking model is aligned to the public band descriptors: it extracts the measurable features above from your recording, maps them to the four criteria and reports an estimate per criterion with the evidence behind it, such as the pauses it counted or the words it flagged. The estimate is deliberately presented as practice guidance, close enough to show which criterion is holding you back and whether you are trending towards your target, and not as an official result. For a verified band, a correction reviewed by a real examiner can be added on top. The AI feedback page describes this two-layer model and what the AI does and does not claim.

Using it: record, review, re-record

1. Record a full Part 2 talk or three Part 3 answers under the real timing, without stopping to restart. The point is to capture what you actually do under pressure.
2. Review only the lowest criterion. If Fluency is lowest, listen to the recording at the timestamps where pauses were flagged and notice what you were searching for: a word, a grammar form or an idea. Each has a different fix.
3. Re-record the same answer once, targeting that one thing. Compare the two reports. A ten-second drop in total pause time, or three fewer fillers, is real progress and is visible in a way that "I think I did better" is not.

Repeat daily with new questions. After two weeks, read the trend across recordings rather than any single report; single reports vary with the question, the trend does not.

What a good report looks like

A useful speaking report has the same shape whatever tool produces it. It gives a separate estimate for each of the four criteria rather than one band. It attaches evidence to each: timestamps for pauses, the exact words it heard for pronunciation flags, the repeated items for vocabulary. It distinguishes error patterns from one-off slips, because a sound you mispronounced once is noise and a sound you mispronounced in every answer is a target. And it names one thing to change next. A report that gives a single band and a paragraph of general advice is not feedback you can act on, however accurate the band might be.

Three mistakes people make with AI feedback

• Chasing the number. Re-recording the same answer five times to push the estimate up teaches you that answer, not the skill. One re-record per answer, then move on.
• Ignoring the evidence. The flagged pauses and words are the feedback; the band is a summary. Read the evidence.
• Skipping the human step. Before the test, one examiner-reviewed correction tells you whether the AI estimate matched a human judgement for your voice and accent. That calibration is worth more than another twenty recordings.

Practise with Langujet

Langujet IELTS speaking practice runs all three parts with the real timings, scores each on the four criteria and shows the evidence for each estimate. Use the IELTS Speaking practice tests for full-length runs and the recall-based topic lists in our Part 1, 2 and 3 guides for daily questions. When the trend is stable, add an examiner-reviewed correction before you book.