Is ChatGPT Accurate for IELTS Speaking? What the Research Shows

ChatGPT can’t give you an IELTS Speaking band you can rely on. It can chat with you and correct your grammar, but we could not find a single study that checks its IELTS Speaking bands against IELTS examiners. When we gave ChatGPT the same answer in 20 new chats, its overall estimate stayed at 6.0–6.5, but individual criteria moved by up to a full band, and four times it scored pronunciation from text alone. In a typed chat, or when your voice is turned into text before ChatGPT reads it, it cannot hear your pronunciation, and pronunciation is a quarter of your Speaking band.

PrepEx’s IELTS Speaking practice is built for the test instead. An examiner voice asks the Part 1 and 3 questions, your answers are recorded, and your pronunciation is scored from the audio, with estimated bands for all four criteria. This guide covers what ChatGPT misses, what our test and the research show, and what PrepEx does differently. The studies and official pages it cites were checked on 14 October 2026.

Try PrepEx Speaking practice
Start with a free account: sample Speaking tasks, two IELTS practice sessions a day and estimated bands for all four criteria.

We asked ChatGPT to score the same answer 20 times

On 14 October 2026 we gave ChatGPT one Part 2 answer and asked for IELTS Speaking bands, 20 times. We used a new Temporary Chat for every run, set to Unpersonalized so that memory and custom instructions were ignored, with the default model and Think left off. The interface did not name the model.

The answer is our own, written to read like a transcript of a band 6 candidate. It responds to the cue card about helping a stranger from our free mock test, with everyday vocabulary and several tense and agreement errors. We alternated two versions: the spoken version below, with its “um”s, “uh”s and two restarts, and a cleaned version with only those removed, which is what many speech-to-text tools produce. The prompt asked for a band for each criterion and an overall band.

Um, okay, so I'm going to talk about a time when I helped a woman I didn't know. It was, uh, last winter at the train station in my city. I was waiting for my train to university and I saw a old lady who looked very confused, and she had two big bags. So I, um, I went to her and asked if she need some help. She told me she want to go to her daughter's house, but her phone was dead, so she cannot check the address, and she also didn't know which platform she should go. Uh, so first I helped her with the address. She remembered the name of the street, so I searched it on my phone and, you know, I found the right train for her. Then I carried one of her bags and we walked together to the platform, and I waited with her until the train come. Actually I was a little bit late for my class that day, but I didn't mind. How I felt afterwards... um, I felt really happy and useful, because in my opinion helping other people is very important, especially old people who are not familiar with technology. She said thank you many times and she gave me a, uh, a chocolate. So it was a small thing, but it made my day, and since then I always try to help people when I see they are in trouble.

• The overall estimate was stable. Every reply called its bands estimates, and every table put the overall at 6.0 or 6.0–6.5. In two replies, though, the opening sentence said 6.5 while the table below it said 6.0–6.5.
• The criteria moved. For the same words, fluency ranged from 6.0 to 7.0 and grammar from 5.0 to “5.5–6.0”. The 20 chats produced six different combinations of fluency, vocabulary and grammar bands.
• Pronunciation was sometimes guessed. In 16 chats ChatGPT said it could not score pronunciation from a transcript. In the other 4 it gave a 6.0 anyway, marked as provisional or illustrative.
• Removing hesitations helped fluency, not the overall. With the “um”s and restarts, fluency got 6.0 in three of ten chats. Without them, it never went below 6.5 and once reached 7.0. The overall estimate did not rise with it.

This is one constructed answer and 20 chats with one model on one day, so treat it as an illustration, not a benchmark. No examiner rated the answer, so the test shows how consistent ChatGPT was, not how accurate. The full answer, prompt and every band are in our test record (JSON).

What PrepEx does that ChatGPT can’t

ChatGPT is a general assistant. PrepEx’s IELTS Speaking practice is built for the test: it asks the questions, listens to your answers and scores them against the four IELTS criteria.

• An examiner voice asks the questions. In Parts 1 and 3, the examiner reads each question aloud, asks “Why?” or a follow-up question based on what you said, and repeats the question if it did not hear you. You answer out loud, as you will in the test.
• Your pronunciation is scored from your recording. A speech-assessment model listens to the audio, so your pronunciation band comes from how you actually sounded, not a guess from text. In our test, ChatGPT declined to score pronunciation 16 times and guessed 6.0 four times.
• All four criteria, with evidence. Fluency and coherence, vocabulary and grammar are rated against the public IELTS band descriptors, using your transcript and how long you spoke. The feedback quotes your own words when it explains what to change.
• Replay and try again. Listen back to your answer, record a new attempt and keep your reports in one place, so you can see whether the change you practised worked.
• Stronger versions of your own answer with IELTS Pro. The Examples tab shows a manageable improvement and a stronger complete version of your saved answer, with an explanation.

A free PrepEx account includes sample Speaking tasks and two IELTS practice sessions per UTC day, and the first diagnostic, which includes Speaking, is free. IELTS Pro unlocks all topics and full Speaking tests. Like every practice tool, PrepEx gives estimated bands, not an official IELTS result.

Record an IELTS Speaking answer in PrepEx

Want a person’s judgement too? Book a 30-minute live speaking mock with an IELTS teacher on Google Meet, including feedback. PrepEx runs these sessions; they are practice, not an official IELTS test. To rehearse all three parts with the official timings first, take our free IELTS Speaking mock test.

How is IELTS Speaking scored?

The IELTS Speaking test is an 11–14 minute face-to-face interview with a certificated examiner, and it is recorded (official Speaking format). The examiner rates four criteria of equal weight: fluency and coherence, lexical resource, grammatical range and accuracy, and pronunciation. Your Speaking band is the average of the four (IELTS scoring in detail).

Two of the four criteria depend on sound. IELTS’s key assessment criteria list speech rate and interruptions such as false starts, repetition and pauses to search for words as evidence of fluency. For pronunciation, they list how you divide speech into meaningful chunks, rhythm, the way sounds link together, stress and intonation, individual sounds and word stress, and how much your accent affects how easily you are understood. A written transcript contains none of the pronunciation features, and speech-to-text tools remove most of the fluency evidence, as the next sections show.

Examiners are trained and certificated, and they recertify every two years. IELTS monitors their marking using Speaking test recordings and reports a reliability figure of 0.90 for Speaking when uneven score profiles from 2024–2025 were marked a second time (IELTS test statistics). Human marking is not perfect either, but it is checked. Our IELTS Speaking test guide explains each criterion with a band 6 and band 7 comparison.

Every automated Speaking score is an estimate, including those from the organisations behind IELTS. The British Council’s IELTS Ready gives estimated scores for its mock tests (IELTS Ready). IDP Vietnam’s guide to preparing with AI says predicted bands are for reference only and do not replace examiners’ marking, especially for Speaking and Writing (IDP Vietnam, in Vietnamese). Cambridge says the score from its Speak & Improve tool “is only a guide to help you practise” (Speak & Improve).

What does research say about ChatGPT and speaking scores?

We could not find any study, published or preprint, that compares ChatGPT’s IELTS Speaking bands with scores from IELTS examiners. The closest evidence comes from other speaking tests, from pronunciation tasks and from IELTS Writing.

ChatGPT reading a transcript of speech

A 2024 study from Kobe University gave ChatGPT text transcripts of 140 learners’ recorded conversations, along with 140 essays from the same learner archive, and compared its scores with the average of 80 trained raters. For the essays, GPT-4’s scores correlated with the raters’ at 0.89. For the speech transcripts the figure was about 0.56, and the best prompt raised it to 0.64 (Uchida, 2024). A correlation of 1 would mean ChatGPT’s scores rose and fell exactly with the raters’. The same model judged writing far better than speech it could only read.

GPT-4o listening to the audio

Hearing the recording did not close the gap. Researchers at Microsoft had GPT-4o listen to 2,500 English sentences read aloud by Mandarin-speaking learners and score their pronunciation. Its scores correlated with the experts’ at about 0.45 to 0.50, while specialist pronunciation-assessment systems reached 0.74 to 0.81. GPT-4o did not return a complete score for 41–48% of the recordings, and two identical runs gave different scores (Wang and colleagues, 2025; a preprint).

General models avoid low scores

In a 2025 study at Radboud University, an open audio model scored 5,000 read-aloud sentences on a 0–10 scale without any training for the task. Its scores bunched between 7 and 9. Experts had rated 863 of the sentences 6 or lower for accuracy, and the model gave none of them a score that low (Parikh and colleagues, 2025).

A Cambridge team found a similar pattern. Used straight out of the box, an audio model gave mostly middle grades and correlated with human scores at about 0.64. After training on more than 30,000 scored test responses, the same model reached about 0.94 to 0.95 (Ma and colleagues, 2025). Automated systems can assess speaking when they are built and tested for it. A general chatbot asked “what band is this?” has not been.

These figures come from different tests, learners and methods, so compare results within a study rather than across studies. None of the studies used IELTS bands, and none tested GPT-Live or the newest ChatGPT models. Newer models may do better, but we found no published evidence either way.

Why a transcript can make your speaking look better than it sounded

Many ways of getting your speech into ChatGPT turn it into text first: dictation, transcription apps and ChatGPT’s Standard voice. These tools are designed to produce readable text, so they leave out much of what an examiner hears.

• Hesitations disappear. On real learner speech from a Cambridge English test, a standard Whisper speech-recognition model kept 5 of 2,661 hesitations and none of 358 half-spoken words (Ma and colleagues, 2023). In another study, Whisper’s largest model wrote down 64 of 157 filler words, such as “um”, in learners’ unscripted speech (McGuire, 2025; a preprint).
• Some mistakes are tidied away. In one of the Cambridge examples, a learner said “um”, restarted, and then said “twenty dollar”. The transcript read “$20 per week”, so the hesitation, the false start and the missing plural all vanished.
• Your accent affects the transcript. A study of Whisper found more errors for speakers with less English experience, and for speakers of tone languages such as Mandarin, Vietnamese and Thai (Graham and Roll, 2024). If a word is recognised wrongly, ChatGPT may comment on a mistake you did not make.

OpenAI’s Help Center says ChatGPT’s voice transcripts are not verbatim and may not match exactly what was said (ChatGPT Voice). A transcript is still good for checking your words and grammar. It is weak evidence of fluency and no evidence of pronunciation. Our own test points the same way: ChatGPT gave the cleaned version of our answer a slightly higher fluency band than the version with its hesitations.

Can ChatGPT voice mode hear your pronunciation?

It depends on the mode. OpenAI’s Help Center describes three voice options:

• Standard transcribes your speech before the model replies, so the model works from text.
• Advanced processes audio within a single model.
• Live, the newest option, runs on GPT-Live models that process audio and can listen and speak at the same time (OpenAI).

Dictation also turns your speech into editable text, which is then sent as a normal message (Voice Dictation FAQ). So in a typed chat, with dictation or in Standard voice, ChatGPT cannot hear how you said anything.

Live and Advanced voice do process sound, but hearing is not the same as scoring. OpenAI makes no claim about the accuracy of pronunciation or speaking scores, and the GPT-4o study above found its pronunciation scores well behind specialist tools. OpenAI also trained GPT-4o to perform the same for speakers with different voices and accents (GPT-4o System Card). That helps it understand you. Our reading is that a model built to understand everyone equally well is not built to notice how much effort your pronunciation asks of a listener, which is part of what the examiner rates.

Why ChatGPT gives a different band each time

Ask ChatGPT to score the same answer in two new chats and you can get two different bands. Four things cause most of this.

• Answers vary between runs. In one test, five language models set up to give repeatable answers still varied in accuracy by up to 15% across 10 runs (Atil and colleagues, 2024). GPT-4o’s two identical pronunciation runs in the Microsoft study also differed.
• Small prompt changes move the score. In the Kobe study, adding a teacher persona and then explanations of the criteria changed how closely ChatGPT agreed with the raters. In a Cambridge study of an open language model, changing only the order of the band descriptors in the prompt shifted its scores on three criteria (Bannò and colleagues, 2025).
• Without a fixed rubric, grades drift. In a study of essays from marketing courses in Chile, ChatGPT-4o gave the same answer different grades in different chat sessions. A rubric reduced the variation but did not remove it (García-Varela and colleagues, 2025; writing, not speaking).
• The model changes. OpenAI updates ChatGPT’s models regularly, and in 2025 it rolled back a GPT-4o update because the model had become too flattering and agreeable (OpenAI). A band from last month may not be comparable with one from today.

Our 20-chat test showed this on a small scale: the overall estimate held steady while the criterion bands moved. Consistent scores are not automatically accurate, either. In an IELTS Writing study, ChatGPT scored the same 30 essays very consistently across three sessions, yet about 0.9 of a band below certified IELTS examiners on average (Torki and colleagues, 2025; a preprint). In another IELTS Writing study, ChatGPT’s overall agreement with official Task 2 bands looked strong. Counting essay by essay from the paper’s appendix, it matched the official band exactly for 18 of 55 essays and was a band or more away for 16 (Koraishi, 2024; our count). We looked at the Writing side in more detail in Can AI score IELTS Writing accurately?

Practising IELTS Speaking with ChatGPT: what to watch for

ChatGPT can add extra conversation practice between PrepEx sessions if you keep its limits in mind.

• Speak, don’t type. A typed answer is writing practice, not speaking practice.
• Record yourself, on a second device if ChatGPT is using your phone’s microphone, so you can hear your own fluency and pronunciation.
• Ask for corrections that quote your words, not a band. In our test, the criterion bands changed from chat to chat.
• Don’t memorise its model answers. Examiners are trained to recognise rehearsed responses, and the British Council says these can limit your score (IELTS Speaking preparation).

Our guide to finding an IELTS speaking partner includes an examiner prompt and ways to practise with people.

How to judge any AI IELTS Speaking score

Whichever tool you use, check these five things before you believe a band.

• Did it hear you? If it only received text, its pronunciation band is a guess and its fluency band rests on very little.
• Does it show evidence? Useful feedback quotes your words and explains what to change. A band with no evidence gives you nothing to practise.
• Does it repeat itself? Submit the same recording twice. A large difference means the number is unstable.
• Has it been tested? Look for how many recordings were compared with human examiners and which measure was used. An exact band match, a result within half a band and a correlation are different things.
• Does it promise too much? No practice tool can give you an official IELTS score or guarantee your band on test day.

Get your Speaking band estimate with PrepEx

Sources and editorial note

Every study and official page linked above was opened and checked on 14 October 2026. Preprints have not been peer reviewed and are marked as such. Correlations and match rates come from the papers’ own tables; the Koraishi essay counts are our count from that paper’s appendix. We ran our own ChatGPT test on 14 October 2026; its record is linked in that section. For the IELTS format and criteria, see IELTS’s Speaking test format and the Speaking band descriptors.

PrepEx’s editorial team wrote this guide. PrepEx offers the IELTS Speaking practice with AI feedback and the live speaking mocks described above. PrepEx is independent of OpenAI, the British Council, IDP and Cambridge University Press & Assessment.