How AI scores IELTS Speaking (and where it still gets it wrong)

Speaking · 6 min read · By Ilyas Muhammad

You record a two-minute answer, and a few seconds later an AI tells you "Band 6.5". It feels like magic, but it isn't. Here's what actually happens between your voice and that number, which parts AI does well, and where you should still be sceptical.

Step 1: your voice becomes text

First, a speech-recognition model turns your recording into a transcript. A good one keeps the details an examiner listens for: word-level timestamps, pauses, and fillers like "um" and "you know". A tidy, cleaned-up transcript would hide exactly the hesitations that affect your fluency score, so for scoring you want the messy, verbatim version.

Step 2: the four official criteria

IELTS Speaking is marked on four criteria that count equally. An AI examiner tries to judge each one separately, the way a human examiner does:

• Fluency and Coherence: how smoothly you keep going, how often you pause or restart, and whether your ideas connect logically.
• Lexical Resource: the range and precision of your vocabulary, including natural word partnerships (collocations) and paraphrasing.
• Grammatical Range and Accuracy: whether you use a mix of simple and complex sentences, and how many errors you make.
• Pronunciation: how easy you are to understand, including individual sounds, word stress, and intonation.

A language model reads the transcript against the public band descriptors and gives a band for the first three criteria, with evidence: the sentences that show strong vocabulary, the grammar slips, the places you lost your thread. Those four scores are then combined into your overall Speaking band.

The most useful part of an AI score isn't the number. It's the evidence: "you said more better twice" is something you can fix tonight.

Step 3: pronunciation is the hard part

Fluency, vocabulary and grammar can be judged from a transcript. Pronunciation can't, because the transcript only shows what the model thought you said. Serious pronunciation scoring compares the actual sound of each phoneme with how it is expected to sound, using techniques like forced alignment and "goodness of pronunciation" scores. That's more demanding to run, and it's still the least reliable of the four criteria, especially across very different accents.

Where AI scoring still gets it wrong

• It's an estimate, not an official result. Only a certified IELTS examiner can give you a real band. Treat AI scores as a practice thermometer.
• Accents can confuse transcription. If the speech-recognition step mishears you, every later step inherits that error. Always check the transcript.
• Half-band differences are noise. If two attempts score 6.0 and 6.5, don't celebrate or panic. Look for trends over ten answers, not one.
• It can't see the room. A human examiner reacts to you in real time. AI only sees your answer to one prompt.

How to use AI scores well

1. Record an answer, then read the transcript before looking at the score.
2. Pick the one criterion with the lowest band and fix only that this week.
3. Re-record the same question and compare the evidence, not just the number.
4. Turn your repeated mistakes into flashcards so they stop coming back.

Related: IELTS Speaking Part 2: a 4-step method to ace the cue card.