Put two recordings side by side. The first speaker has a heavy accent — you can hear the vowels leaning, the consonants landing a millimetre off — and you understand every word without effort. The second speaker's accent is lighter, closer to the standard, and you have to work. You ask them to repeat. You lose the thread of the sentence halfway through.

Almost every learner has met both of these people, and almost every learner who has met them has drawn the wrong conclusion from the comparison. The conclusion is usually "I need to sound more native." The evidence, which has been accumulating since a 1995 study in the journal Language Learning, says that is the wrong target — and that a great deal of what learners drill is aimed at the wrong part of the problem.

Three Dimensions, One Confusion

In 1995, Murray Munro and Tracey Derwing published an experiment with second-language learners that measured three things about the same set of recordings and found that they did not move together. The three things have been the core vocabulary of the field ever since:

  • Intelligibility — "the extent to which a speaker's message is actually understood by a listener." Operationalised in the lab by asking listeners to transcribe what they heard. Either the words arrived or they did not.
  • Comprehensibility — "listeners' perceptions of difficulty in understanding particular utterances." Measured with a rating scale: how much effort did that cost me?
  • Accentedness — how closely the pronunciation approaches a native speaker's. Measured on an accent-strength scale. This is what learners mean when they say "my pronunciation is bad."

The finding that made the study famous was a negative one. A strong foreign accent did not necessarily reduce comprehensibility or intelligibility. In their own words, twenty-five years later:

"The key finding was that the dimensions at issue are related, but partially independent. Of particular note was our observation that speech can be heavily accented but highly intelligible." — Munro & Derwing, "Foreign accent, comprehensibility and intelligibility, redux," Journal of Second Language Pronunciation 6(3), 2020

That left an obvious question: how independent, exactly? A meta-analysis published in Studies in Second Language Acquisition in 2025 (Chau & Huensch) analysed 49 reports from 1995 to 2023, covering 141 effect sizes, and produced a table of weighted mean correlations that is worth reading twice:

  • fluency and comprehensibility: r = .82
  • comprehensibility and accentedness: r = .75
  • fluency and accentedness: r = .62
  • intelligibility and comprehensibility: r = .57
  • intelligibility and accentedness: r = .32

Look at the bottom row. The relationship between how accented you sound and whether your meaning actually arrives is the weakest link in the entire network — r = .32, or roughly ten per cent of the variance. The two dimensions learners treat as a single thing are the two most loosely connected things in the field. The meta-analysis drew the pedagogical conclusion plainly: having an accent does not mean being unintelligible, which supports prioritising intelligible and comprehensible speech over accent reduction.

What Listeners Actually Use

The second meta-analysis to know is Kazuya Saito's, which pulled together 37 listener studies and 17 training studies and asked what listeners are actually reacting to:

  • Accentedness judgements were strongly tied to consonants and vowels — segmental accuracy. If you want to sound native, the individual sounds are what listeners are grading. (This is also the part instruction struggles to move; more on that below.)
  • Comprehensibility judgements were related to a range of segmental, prosodic and temporal features — not just sounds, but word and sentence stress, rhythm, speaking rate and pausing. Comprehensibility is a broader, more forgiving construct. You have more levers than you think.
  • Expert listeners leaned more on phonological information than untrained ones. It matters who is judging you.

Saito's formulation is the useful one: comprehensibility and accentedness are "readily distinguishable" because raters rely on different kinds of information to make the two judgements. Two questions, two different scorecards.

Not All Errors Cost the Same: The Functional Load Principle

If intelligibility is the target, the next question is which of your errors are sabotaging it. Linguistics has an answer that predates the empirical test by decades: functional load, defined by Harold King in 1967 as "a measure of the work which two phonemes do in keeping utterances apart."

Some sound contrasts carry a lot of traffic. The /l/–/n/ distinction separates many common word pairs; /l/ and /n/ appear in minimal pairs at the beginnings and ends of words all over the language. Other contrasts barely work at all: the /ð/–/d/ opposition distinguishes very few pairs, and plenty of native English speakers do not observe it. On paper, then, an error on /l/–/n/ should cost a listener far more than an error on /ð/–/d/.

Munro and Derwing tested that in 2006 with 13 native English listeners judging 23 Cantonese-accented sentences containing deliberate combinations of high and low functional load errors. The results:

  • High functional load errors had relatively large effects on both accentedness and comprehensibility ratings.
  • Low functional load errors had only a minimal impact on comprehensibility — listeners barely noticed them.
  • The only cumulative effect in the data appeared with multiple high load errors, and only on accentedness judgements.

Later work moved the finding out of the lab and into real learners. A study with 40 Japanese learners of English in an EFL setting found that only high functional load consonant substitutions — the textbook examples being /l/ pronounced as /r/, and /v/ as /b/ — actually damaged native listeners' comprehensibility judgements. Low load substitutions did not. A 2023 partial replication under English-as-a-lingua-franca conditions (24 sentences, 171 Japanese listeners transcribing target words) confirmed the pattern, and found that several high load substitutions in a single word depressed intelligibility further than one.

The Lingua Franca Core: What Survived the Corpus

The most influential attempt to build a pronunciation syllabus out of intelligibility data is Jennifer Jenkins's Lingua Franca Core, drawn from recordings of non-native speakers talking to each other. Features that caused breakdowns went in the core; features that never seemed to cause trouble were left out. Roughly, the core includes:

  • most consonant contrasts, with the standard exceptions of the th sounds (/θ/ and /ð/) and the dark or velarised /l/, which may be substituted without costing understanding
  • initial consonant clusters — no dropping or inserting extra vowels inside them
  • vowel length contrasts, such as the difference between the vowels of "live" and "leave"
  • nuclear or tonic stress — the word that carries the point of the sentence

Non-core, on this account: exact vowel quality, word stress, the weak forms and reduced rhythm that give British English its texture. Note what that leaves out. Word stress and the whole "typical rhythm" of a native accent are precisely where learners spend enormous effort.

Two cautions, because this is a working model rather than a settled law. Jenkins herself stressed that the Core is not a model for imitation — it is a list of features whose absence causes miscommunication, and the sensible target is the local teacher's accent containing the core features. And later research has pushed back at the edges: studies by David Deterding found some non-core features doing intelligibility work after all, and experiments manipulating word stress, consonant clusters and vowel length found that word stress — a non-core item — did affect lingua franca intelligibility. Treat the Core as a well-founded priority list, not scripture.

The Asymmetry That Should Change Your Whole Plan

Here is the finding that should reorganise how you study, because it separates the two goals in time.

Derwing and Munro followed adult immigrant learners of English for seven years, beginning shortly after their arrival in Canada and continuing to year ten. The participants were 25 first-language Mandarin and 25 Slavic-language speakers, all beginners at the start, all enrolled in ESL classes but none receiving systematic pronunciation instruction. Their speech was judged at three points by two sets of listeners: native monolinguals and highly proficient L2 speakers of English. The published result:

"The Mandarin L1 speakers showed no change over time on any of the dimensions, while the Slavic language L1 speakers improved significantly in comprehensibility and fluency. Improvement in accent was limited to the first 2 years in the Slavic language group." — Derwing & Munro (2013), Language Learning 63(2), 163–185

Read the middle clause again. In seven years, with the learners living and working in English the whole time, accent stopped moving after year two, while comprehensibility and fluency kept improving. Related measurements from the same project put the largest vowel intelligibility gains in the first six to eight months for the Slavic group and the first eight months for the Mandarin group — a "window of maximal opportunity" that opens on arrival and narrows fast.

The Mandarin group's flat line is not a comfort either, and the researchers' own explanation is actionable: the difference appears to lie in willingness to communicate — how much of the learners' lives happened in English, and how much they talked. Pronunciation development follows use, and the naturalistic data says it is front-loaded. The window is early; the work is not over when the window closes, because comprehensibility responded to instruction even in learners who had been in the country for years.

Instruction Works — If You Aim It at the Right Target

The clearest experiment on where to aim is Derwing, Munro and Wiebe's 1998 comparison. Learners of English as a second language were split into three groups: one got segmental instruction (individual sound contrasts), one got global instruction (speaking rate, intonation, rhythm, projection, word and sentence stress), one got no pronunciation-specific instruction at all.

The segmental group did improve on measured sound accuracy — a detailed error analysis showed real gains in the individual sounds. Those gains did not show up in listeners' overall ratings. The global group's gains did: it improved significantly on comprehensibility and fluency. Instruction on the parts of speech that carry meaning across a sentence outperformed instruction on the parts that make it sound native.

Saito's meta-analysis of 17 training studies generalises the point: instruction produced larger gains in comprehensibility than in nativelikeness, and the advantage became clearer when the treatment targeted prosodic accuracy. That meta-analysis ends with a blunt sentence practitioners should tape to their wall: attaining native-like pronunciation may require an exclusive focus on segmental refinement, which is resistant to the influence of instruction.

The Other Half of Intelligibility Is Your Listener

Everything above treats intelligibility as a property of your mouth. It is also a property of your listener's ears and experience, and the research is clear about that too.

Bent and Bradlow's 2003 study recorded native speakers of Chinese, Korean and English and played the sentences to native English listeners and to Chinese and Korean listeners. For the native English listeners, the native English talker was most intelligible — as expected. For the non-native listeners, a high-proficiency non-native talker from their own language background was as intelligible as the native speaker. They called it the matched interlanguage speech intelligibility benefit. Later work complicates the picture: one replication did not confirm even the benefit for same-language listeners, and a follow-up study using a word-recognition task found a mismatched detriment for low-proficiency speakers — meaning that two relatively inexperienced non-native speakers who share no first language can genuinely struggle with each other. A study using reaction time as a more sensitive measure found non-native listeners responding faster to same-language speakers, while native listeners were consistently fastest with native speech.

Then there is adaptation. Reaction-time research shows that the cost of processing an unfamiliar accent is real but shrinks fast — the difference diminishes within about a minute of exposure to the same voice. Your first sentence may cost a listener effort your tenth sentence does not.

One more strand, because it circulates widely and deserves the full picture. Lev-Ari and Keysar's 2010 study found that trivia statements were rated less truthful when spoken with a heavy accent (a mean of 6.9 on the truthfulness scale versus 7.47 for a native accent). Their key nuance gets dropped in retellings: a mild accent showed no penalty at all (7.52, statistically indistinguishable from native). The effect tracked difficulty of understanding, not the presence of an accent. And the finding is contested — a direct replication in 2013 failed to reproduce it, a 2021 follow-up did reproduce it, and a 2024 conceptual replication found no credibility drop and even higher trustworthiness ratings for the foreign-accented speaker. The honest summary: the phenomenon is real enough to be studied seriously and unstable enough that you should not build your self-image on it. Either way, the trigger is comprehensibility. When you are hard to process, everything you say gets a little less persuasive — and that is a target you can train.

What to Train Instead

1. Diagnose with a transcription test, not a feeling

Record sixty seconds of yourself answering a real question, unscripted, and send it to three people who are not you and not your teacher. Ask them to transcribe what you said. The intelligibility measure in the research is transcription; use the same one. Score it two ways: what percentage of words arrived, and which words did not. Then ask them separately how much effort it took, on a one-to-five scale — that is the comprehensibility rating. Two numbers, and a list of the words that failed. Repeat every four weeks with a new recording, and keep the old ones, because your memory of how you sound is useless as a baseline.

2. Rank your errors by functional load and drill the top of the list

Take the failing words from your transcriptions and sort them by the contrast involved. The high-load contrasts to prioritise: /l/ versus /n/, /l/ versus /r/, /p/ versus /b/, /t/ versus /d/, /f/ versus /v/, /s/ versus /z/, and vowel length pairs such as /iː/ versus /ɪ/ and /uː/ versus /ʊ/. Put the th sounds and the dark /l/ at the bottom of the pile; the research says listeners absorb those substitutions without much trouble. A minimal-pair deck built only from high-load contrasts will do more for you than a general "pronunciation" deck, because it is aimed at the errors that change what people hear.

3. Then spend your remaining effort on prosody, not on the last few vowels

Word stress, sentence stress, rhythm and chunking are where instruction reliably pays, because they are where comprehensibility lives. Cards that carry a stress shift are ideal for this: REcord versus reCORD, PREsent versus preSENT, EXport versus exPORT. For sentences, use nuclear stress — the practice of deciding which single word carries the point — and let the rest of the sentence reduce. This is the target Jenkins put in the core and the one Derwing and Munro's global instruction group improved on while the segmental group's measured gains stayed invisible to listeners.

4. Control tempo

Comprehensibility ratings track articulation rate and pause behaviour. Faster is not better; steady is better. When you practise a hard sentence, practise it at the speed where every word lands, then keep that pace in real conversation rather than accelerating under pressure. Pausing in the right places beats pausing less.

5. Keep a miscommunication log

This is the highest-signal record you can keep. Every time someone asks you to repeat, misunderstands a booking, spells your name wrong or answers the wrong question, write down the sentence and the word that broke. After a month, sort it. You will usually find twenty per cent of the entries are one or two recurring items — a vowel length, a word stress, a cluster at the start of a word. That is your curriculum, written by your actual listeners rather than by a textbook's order of presentation.

6. Rehearse the listener's side by listening to more accents

Intelligibility is a two-way skill, and the listener research cuts both ways: your ear benefits from the same variability that helps everyone. Listening across many voices, not one, is the training regimen that generalises to new speakers. If you spend two years with one podcast host, you have trained for one listener.

What to Stop Doing

  • Chasing a native accent as the goal. It is the dimension with the weakest relationship to being understood (r = .32), and the seven-year data suggests it largely stops responding after about two years of naturalistic exposure. Set your target at comprehensible and intelligible in the specific situations your life actually contains, and treat accent as whatever remains.
  • Fixing th first. It is the poster child for a low functional load error and one of the substitutions listeners handle most easily. It is also the error learners notice most in their own recordings, which is exactly why they over-train it.
  • Judging yourself from your own recording. You are the worst available judge of your own accent, because you know what you meant. Get transcriptions and effort ratings from other people. That is the measurement.
  • Assuming your accent quality is a fixed verdict. The Slavic group in the seven-year study kept getting more comprehensible and more fluent long after accent had stopped changing, and pronunciation responded to instruction in learners that models of "fossilization" had written off. Accent may settle. Comprehensibility does not have to.

The Bottom Line

The research over three decades has done something useful, and slightly uncomfortable: it has separated the thing learners work hardest on from the thing that decides whether they are understood. Accent is real, it is measurable, it correlates with intelligibility at r = .32, and after a couple of years it often stops moving. Comprehensibility responds to prosody, tempo, high-load sound contrasts and willingness to open your mouth, and it keeps responding for years.

So spend your study time where listeners actually lose you: on the handful of contrasts that carry real information, on the stress pattern that tells them what you meant, on the tempo that gives them time, and on the conversations that pull all of it into use. Your accent is a record of where you learned the language. Your intelligibility is a skill, and it is one you can move with data — sixty seconds of recording, three honest transcriptions, one log, once a month.