Does your accent affect your IELTS Speaking score?
Examiners mark intelligibility, not nationality. What the pronunciation criterion actually measures, and why speech apps flag things examiners ignore.
Band descriptorsSpeakingUpdated 8 min read
No, your accent is not costing you marks. Pronunciation is scored on two things — how much effort a listener spends understanding you, and the range of pronunciation features you control — and accent is neither of them; the public descriptors accept first-language influence at the top bands provided it rarely interferes with understanding. That criterion is unpacked band by band in what the Pronunciation criterion measures. This article answers the narrower question that sends candidates down a years-long detour: why does automated feedback keep rejecting pronunciation an examiner would accept, and how do you work out which of the two is right about any particular word?
The pattern is familiar. You say a word cleanly, the same way twice, and the app marks it wrong, because it is comparing your sounds against one reference accent that happens not to be yours. That is a limitation of the tool, not a diagnosis of your English.
Where a pronunciation score and an examiner part company
The distinction worth internalising is between features that change whether a listener understands you, and features that only change whether you sound like a particular region. The descriptors care about the first column. Automated scoring, which has to work by matching acoustic patterns to a model, flags both.
| What the Pronunciation criterion rewards | What automated scoring flags and examiners do not |
|---|---|
| Word stress in the right place: acaDEMic, not aCAdemic. Listeners retrieve words by stress shape, so this is the highest-value fix available. | Vowel colour: your a in dance, your o in thought. These vary across the English-speaking world, and neither word has a rival that the variation could turn it into. |
| Sentence stress: content words carrying the beat, function words compressed. The RENT was HIGHer than I exPECted. | Rhoticity: whether you pronounce the r in car or hard. Both are standard, and both are heard in test centres. |
| Weak forms: It should have been finished by now, with should have been run together and unstressed rather than delivered as three equal, fully voiced syllables. | Regional consonants with no rival word behind them: a glottal t in butter, a tapped r in very, a dark or clear l in milk. In these particular words, nothing else is a candidate. |
| Chunking: pausing at meaning boundaries, not mid-phrase. In my country, // most graduates // move to the capital. | Pitch range measured against a model speaker. Intonation counts when it carries meaning, not when it matches someone else's melody. |
| Intonation that does work: contrast marked by pitch, lists that rise then fall, a genuine question shaped as one. | Speaking rate against an average. There is no target words-per-minute in the descriptors, and speed usually damages the stress contrasts in the left column. |
| Individual sounds, but only where they collide with another real word: bad heard as bat, closed as close, live as leave. | Phoneme-by-phoneme match with a single reference accent, repeated on the same word however clearly you produce it. |
Read the right-hand column again and notice what it has in common: none of those examples makes a listener guess, and none of them is in the descriptors. The left-hand column is what the criterion is built from, and most of it is not about individual sounds at all. It is about stress, rhythm and phrasing — which is why a candidate can pronounce every vowel carefully, deliver every syllable with identical weight, and still be tiring to follow. The descriptors say only that a band 6 speaker can generally be understood, with mispronunciation reducing clarity at times; equal-weight delivery is precisely the kind of thing that puts the work back on the listener.
The overlap test: your listener's list against the machine's
Since the criterion is listener effort, measure listener effort — and measure the machine on the same recording, so you can see exactly where the two disagree. This takes about twenty minutes and one other person.
- Pick any Part 2 cue card. Record 60 seconds on your phone, one take, no script, no second attempt. Everything below uses this one file.
- Recruit a listener who does not share your first language and does not know your topic. A colleague, a classmate from another country, a tutor.
- Give them one instruction only: while listening once, write down every word you had to guess at and nothing else. No general impressions, no advice, no comments about accent. Guessed words only. This is your human list.
- Now run the same file through any plain dictation or speech-to-text tool — transcription, not pronunciation scoring — and write down every word the transcript got wrong. This is your machine list.
- Compare the two lists word by word and sort each flagged word into one of three places:
| Where the word appears | What it means | What to do with it |
|---|---|---|
| Both lists | Neither a human nor a machine could recover the word from the sound you made. This is intelligibility in the sense the descriptors mean it. | Top priority. Fix these first. |
| Human list only | A transcription engine's language model often repairs a garbled word from the surrounding words; a listener hearing you live may not, and your listener did not. | Real, second priority. |
| Machine list only | The word reached a human intact. The machine's model simply expected a different accent shape. | Ignore it. |
- For each word you are keeping, ask which of three things went wrong: was the stress on the wrong syllable, did the sounds collide with a different real word, or was it swallowed inside a rushed chunk? Write the cause next to the word.
- Read the human list count as your headline number. Zero to two guessed words in 60 seconds means intelligibility is not your ceiling, and your preparation belongs in Fluency and Coherence, Lexical Resource or Grammatical Range and Accuracy instead — what band 7 fluency actually sounds like is the natural next check. Three or four means one or two systematic features are leaking and are worth a targeted fix, but not a rebuilt study plan. Five or more means intelligibility is the first thing to work on.
- Repeat with a different card and a different listener a week later. Keep only the causes that appear both times. Those are your systematic features; the one-offs are noise.
Two takes a week apart with the same cause on both human lists is a finding you can act on. One app rejecting a word that your listener understood without hesitation is not.
What to do with the time you just saved
Accent reduction is a multi-year undertaking with no marks attached. The features in the left-hand column are different: word stress, weak forms and chunking can be learned in a few weeks of deliberate work, and all three carry marks. Learning a feature is not the same as banking a higher band, so be honest about the pace of the band itself: across the four criteria, moving half a band takes something in the order of 200 to 240 hours of instruction and practice on average — how many hours IELTS preparation takes turns that rate into a plan — and it slows down above band 6. That is exactly why the hours should go where the descriptors actually pay.