In 1995, Murray Munro and Tracey Derwing played recordings of 10 Mandarin speakers to 18 native English listeners and asked them to do two jobs. Write down every word they heard. Then rate, on scales of 1 to 9, how easy each speaker was to follow and how foreign they sounded.
The transcripts came back highly accurate. The accent ratings were all over the place, with "a noteworthy proportion of scores at the 'heavily-accented' end of the scale." Same speakers, same sentences. Listeners heard a strong accent and still caught the words.
If you've spent years on your accent and people still say "sorry, can you say that again?", that study points to the likely reason. You've probably been fixing the wrong list. Below: accent vs intelligibility, or the difference between sounding foreign and being hard to follow, the errors that cost you comprehension, why th goes to the bottom of the list, and a 60-second test that tells you which errors are yours.
What's the difference between accent and intelligibility?
Munro and Derwing measured three things, and the names matter because most accent advice treats them as one.
- Accentedness: how far your speech "differs in some noticeable respects from native speaker norms." A listener's rating.
- Comprehensibility: how hard the listener feels they had to work to follow you. Also a rating.
- Intelligibility: how much they actually understood, measured with a transcript scored on exact word matches.
Intelligibility decides whether the pharmacist hands you the right prescription; comprehensibility decides how tired they are by the time you've finished asking for it.
Twenty-five years later the same two researchers looked back on the study and summarized it in one line: the three are "related, but partially independent," and "speech can be heavily accented but highly intelligible."
Related matters too. Accent isn't free, and it's fast: the Bristol team behind a 2015 study of learner speech note that untrained listeners can spot a non-native speaker from "speech that is just 0.03 seconds long." Noticing an accent is not the same as failing to understand it. But the claim the research supports is narrower and more useful: much of what makes you sound foreign isn't what makes you hard to understand, and you can tell the two apart.
Which English pronunciation mistakes actually matter most?
Say need as "leed" and you've produced a different English word, lead. Say mouth as "mouf" and you've produced no word at all. The listener has nothing to confuse it with, so they quietly repair it and move on.
The idea behind that difference is called functional load. A pair of sounds carries a high load when English uses it to tell a lot of words apart, and a low load when it barely does. Swap a high-load pair and the odds are good you land on a real word; swap a low-load pair and you usually land nowhere.
Munro and Derwing tested it in 2006, in what they called an exploratory study of Cantonese speakers, by asking listeners how hard each sentence was to follow. As a teachers' summary of the paper quotes it, "even sentences that contained only one high functional load error were rated significantly worse for comprehensibility than sentences containing three low FL errors." One error beat three.
Yui Suzukida and Kazuya Saito repeated the test with 40 Japanese learners, twice, and got the same answer: only high-load consonant swaps such as /r/ for /l/ or /b/ for /v/ dragged comprehensibility down. Low-load swaps such as "sree" for three or "ze" for the didn't measurably change how hard listeners had to work.
English uses /n/–/l/ and /r/–/l/ to separate a lot of everyday words, so a slip often lands on one. It uses th far less, so most slips land nowhere and the listener repairs them and keeps going.
Adam Brown, who built one of the two standard rankings, tells the best story about it. On a camping trip in Thailand with a group of Singaporeans, his companions kept referring to his "[peʔ]," and Brown had to keep asking for clarification, never sure whether they meant his bag, back, pack or peg. Three high-load contrasts collapsed into one syllable, and the word became four words.
One caution. Functional load is a rule of thumb: a 2021 review notes there is still "no agreed-upon way to define or measure FL." Where the rankings agree with a second, independent shortlist, from the phonologist Jennifer Jenkins, you can trust them more. They agree on one vowel, below, and Jenkins adds one consonant rule of her own.
Leave or live? The one vowel contrast on every list
"When did he leave?" asks for a time. "When did he live?" asks for a century. Brown quotes that pair from a 1990 pronunciation textbook, along with sleep/slip and team/Tim, where students answer each question, so a vowel length slip gets a different, perfectly sensible reply.
The /iː/ of seat against the /ɪ/ of sit is high load in both standard rankings, and it's the only vowel contrast Jenkins kept on her shortlist of must-have features. In her own words, "the contrast between long and short vowels is important."
Both are sensible English sentences, so the listener can't repair it from the words alone.
Context rescues a lot of these. Brown challenges his readers to find one situation where "he wants a ship for his birthday" and "he wants a sheep for his birthday" are both plausible. "I'm living now, see you tomorrow" gets repaired before you've finished the sentence. The damage happens when both readings make sense, as in the line above.
If your first language has one vowel where English has two, there are specific fixes for Spanish speakers and Korean speakers.
String, not sting: consonant clusters at the front of a word
Jenkins is blunt about this one: string "cannot be simplified to 'sting' or 'tring'." Drop a consonant from the start of a word and you've often made another word. Clusters at the end of a word cause a different problem, because that's where English hides its past tense and plurals; the Mandarin speakers' guide to consonant clusters covers how "walked" turning into "walk" makes you sound like a beginner.
Do you need to fix the th sound?
"All the consonants are important except for 'th' sounds as in 'thin' and 'this'." That's Jenkins, writing for teachers in 2002, and th is the first item on her list of features you can leave alone.
The functional load numbers agree. Mouth as "mouf" was Munro and Derwing's example of a low-load error, and Suzukida and Saito classed th said as /s/ or /z/ as low load too. A 2022 study of Turkish speakers had 33 native listeners transcribe target words and found that errors on /θ/ and on schwa "seemed to make speech more accented while not seriously affecting intelligibility."
So "mouf" and "ze" mostly cost you accent points, and by the same logic so do "dis" and "dese". Listeners hear them and understand you anyway. Some th slips do land on real words, think as "sink" or three as "tree", but there are few enough of those pairs that context usually covers you.
That's a reason to fix th last, not never. Jenkins built her list from conversations between non-native speakers, and other researchers have challenged her treatment of th. If the rest of your list is clear and you want the sound, the French speakers' th guide has the tongue drill. Just don't spend your first six months on it.
Try it in Conversa
Practice with AI characters who adapt to your level and give real-time feedback.
Try Conversa FreeStress: the error where every sound is right
"I didn't say he stole the money." English teachers have used that sentence for years because it has seven meanings, one for each word you can stress.
The sounds stay the same in all three. Only the stressed word moves, and the meaning moves with it.
Jenkins calls this nuclear stress and lists it as essential: "My SON uses a computer" implies somebody else doesn't. Munro and Derwing noticed something similar in their own data, writing that prosodic errors "appear to be a more potent force in the loss of intelligibility than phonetic errors." That was one small group of Mandarin speakers, so treat it as a strong hint.
Word stress is where the experts split. Jenkins put it on her leave-alone list. John Field tested it on native listeners, which is who you're mostly talking to if you live or work in an English-speaking country, and on non-native listeners too. Misplaced stress hurt both groups' transcriptions in much the same way. As Mahsa Ghosh and John Levis sum up his results, stress pushed toward the end of the word did more damage than stress pushed toward the front.
You can see word stress doing work in the dictionary: REcord (the noun) and reCORD (the verb) get different pronunciations in Cambridge. And comfortable said as "com-for-TA-ble" pushes the stress right, the direction Field found most damaging, which can leave a listener searching for a word that isn't there. Word stress has its own rules and its own traps, laid out in the guide to PHOto and phoTOgraphy; the stress on thirteen and thirty is a separate trap that catches native speakers too.
Why do people not understand me when my pronunciation is fine?
Dustin Crowther, Pavel Trofimovich, Kazuya Saito and Talia Isaacs compared 45 speakers of Chinese, Hindi and Farsi, and for the Hindi speakers, pronunciation wasn't the main thing slowing listeners down. What hurt comprehensibility depended on the first language.
For the Chinese speakers it was mostly sounds. For the Hindi speakers it was word choice and grammar. The Farsi speakers showed no single cause.
The Hindi speakers still sounded accented, because vowel and consonant errors made everyone sound accented. It just wasn't the accent that made listeners work. If your sounds are clean and people still hesitate, the problem may be a word you're using slightly wrong. "I'm boring" when you mean "I'm bored" is pronounced perfectly and still makes a listener stop and replay what you said.
The 60-second transcription test
Here's the test, adapted from the method Munro and Derwing used:
- Record yourself talking for one minute without a script. Describe your morning, or the last meal you cooked.
- Write down what you said, word for word, from the recording.
- Give the audio to someone who hasn't heard you talk much and ask them to type every word they hear, exactly.
- Compare the two transcripts. Every word they got wrong goes on your list.
- Then ask two quick questions. How hard was that to follow, 1 to 9? How strong is the accent, 1 to 9?
The mismatches are your intelligibility problems. If they typed live where you said leave, your item is the /iː/–/ɪ/ vowel. A perfect transcript with a high effort score can point somewhere less obvious: stress, word choice, grammar, or just pace. And anything they transcribed perfectly and found easy, but rated as strongly accented, is accent. You're allowed to stop paying for it.
One friend and one minute isn't a lab study, and a listener who knows you will probably understand more than a stranger would, so use a fresh listener if you can.
Then pick one item. Kazuya Saito and Luke Plonsky reviewed 77 studies and found pronunciation teaching works best aimed at specific features, practiced while you're paying attention to them. "My accent" is not a feature. Leave against live is.
Be honest with yourself about tools. A 2024 review of speech-recognition pronunciation practice found it much stronger on individual sounds (an effect size of g = 0.82) than on rhythm and intonation (g = 0.37). So an app that scores your vowels is a good fit for leave and live, and a weaker one for sentence stress. The same review found practice under five weeks produced almost no effect (g = 0.07).
If you want somewhere to say your sentences out loud and get a tip on each line, Conversa is an AI conversation partner built for that.
Can you lose your accent, and should you?
Saito and Plonsky's 77 studies show pronunciation does move with targeted practice, so the honest answer to "can you?" is "partly". The better question is what you're measuring. The transcript is the number that decides whether your order comes out right.
Run the test again in five weeks with a new listener, describing the same morning. If the mismatch count dropped, you're probably more understandable than you were, whatever the accent score says.