Day 2 · Linguistics

The Invisible Choreography in Your Mouth

Speaking is not producing sound. It is putting checkpoints in the path of a stream of air — with timing accurate to a few dozen milliseconds.

2026-08-29

How many milliseconds separate your vocal folds starting to buzz from your lips coming apart is something you have never once noticed. Yet that gap decides which word a listener hears, and it is part of why your accent will not budge.

When your nose is blocked by a cold and you say 妈妈 (māma, "mama"), it slides toward something like 巴巴 (bāba). You do not have to wait for a cold to try it: pinch your nose, say it, and you will hear both m sounds collapse on the spot.

That small thing gives away something uncomfortable. The instruction in your head was "say mama," but what actually ran was a whole routine of movements you cannot feel: whether a door at the back of your throat opened or closed, when your vocal folds started vibrating, how far from your teeth your tongue stopped. Pinching your nose only plugged the exit for one of those movements, and the word came out different.

What follows takes the routine apart: where the airflow gets blocked and how; how a difference of a few dozen milliseconds turns into two different words; and why, once this system is tuned, it barely moves in adulthood.

// 01

Speaking is not sound-making, it is blocking air with your mouth

The raw material of speech is a stream of air. The lungs push it up, it passes the two vocal folds in your larynx, and it leaves through the mouth or the nose. What we call "articulation" is mostly not the production of sound; it is putting checkpoints along that path: where to block, whether to block completely or leave a narrow gap, and for how long. Nearly every distinction you hear is some arrangement of those checkpoints.

Start with one you can feel directly. Say 爸 (bà, "dad") and 妈 (mā, "mother"). The lips do exactly the same thing: close, then open. The difference lies in the soft flesh above your throat — the velum (push your tongue tip up and back until the hard roof of your mouth suddenly turns soft; that is it). It works as a door. Raised against the back wall, the route to the nasal cavity is sealed, all the air leaves through the mouth, and you get 爸 (bà). Lowered, the nasal route opens and you get 妈 (mā). When you pinch your nose you do not move the velum at all; you plug the far end of the nasal route, so the air has nowhere else to go but the mouth — which is why 妈妈 (māma, "mama") slumps toward 巴巴 (bāba).

The second dimension is how you block. Put your palm in front of your mouth and try three kinds. Block completely, then release suddenly, and the air bursts out: that is a plosive — b, p, d, t, g, k. Leave a narrow slit and let the air grind through: that is a fricative — f, s, sh. Fricatives can be held: an s can be dragged out until you run out of breath, while the burst in a plosive cannot be stretched at all. Block first and then release into a slit, and you get the two joined together, as in the Mandarin sounds written j, q, z and c.

The third dimension is where you block: the two lips (b, p, m), the tongue tip at the ridge behind the teeth (d, t, n), the tongue body against the hard palate (Mandarin j, q, x), the back of the tongue against the velum (g, k). Say those groups slowly and you can feel the contact point retreating backward.

Air comes up from below, passes the glottis, and gets blocked at some checkpoint in the mouth. The velum is the door that decides whether it leaves through the mouth or the nose — the entire difference between 爸 (bà, "dad") and 妈 (mā, "mother") sits on that one door.

Treat those three as three dials, and every language picks its own set of points on them. Hindi divides the tongue-tip region far more finely than Mandarin does: tongue tip against the back of the front teeth is one place, tongue tip curled up to the front of the hard palate is another, and each of those places carries a four-way split — unaspirated voiceless, aspirated voiceless, voiced, and a breathy-voiced kind. That one region alone yields eight mutually distinct consonants, where Mandarin at the same place has two.

This is not a matter of one language being finer than another. Positions on the dials are continuous, and each language draws its own grid over them; a language with more cells does not thereby say things more precisely, and one with fewer is not short of expressive power. What causes trouble is when the positions of the cells fail to line up: a line your native language never drew is a line you will not go looking for when you listen.

// 02

A few dozen milliseconds is the whole difference between two words

The last section left a question open: English b and p are written with the same letters as Mandarin b and p, so are they the same thing? Not quite. And the difference falls on a quantity you have never consciously registered — time.

Hold your palm in front of your mouth and say 趴 (pā, "to lie face down"): a distinct puff hits your hand. Say 八 (bā, "eight"): almost nothing. What is that puff? It is the stretch during which the lips have already opened but the vocal folds have not yet started vibrating, so air rushes straight out through an open glottis without being turned into voice. Once the folds start up, what leaves is voice rather than air.

So the quantity that matters is the gap between two events: the moment the lips release and the moment voicing begins — which comes first, and by how much. Phoneticians call it voice onset time. It is a signed number: positive means lips first and voice after, negative means the vocal folds start buzzing before the lips let go.

Mandarin sits here: the unaspirated b, d, g run about +10 to +25 milliseconds, and the aspirated p, t, k about +90 to +110. Both groups release first and voice later; they differ by nearly an order of magnitude in how long the wait is. And here is a common confusion worth clearing up: Mandarin has no truly voiced stops at all. What feels like "b is voiced, p is voiceless" is in fact "unaspirated versus aspirated." English b is roughly at zero and English p around +50 to +60 — so the two systems line up reasonably well, with both boundaries landing near +25 milliseconds.

One time axis, three languages drawing the line in different places. Spanish puts its boundary near zero, English near +25 milliseconds — which drops the Spanish p squarely into the English listener's b cell.

Spanish pulls the number over to the negative side. Its b has the vocal folds starting first, buzzing for several dozen milliseconds before the lips come apart; its p sits at zero to ten milliseconds, which is where English keeps its b. The consequence is direct: when a Spanish speaker says the English word "pat", English listeners often hear "bat". Calling that a "strong accent" explains nothing. Two boundaries are drawn at different ticks on one axis, and each person cuts with their own.

Lisker and Abramson measured this quantity for word-initial stops across eleven languages in 1964, and concluded that the great majority of the contrasts those languages make at a single place of articulation can be sorted out by this one timing measure. Strength of evidence: cross-linguistic acoustic measurement over eleven languages, repeatedly confirmed on many more since — one of the sturdiest results in the field.

// 03

Thirty-two years in the country, and you can still hear it

In 1995 Flege, Munro and MacKay recruited 240 Italian-born immigrants in Canada whose age of first learning English ranged from 2 to 23, and who by the time of testing had lived in Canada for an average of 32 years. Native speakers rated their English sentences for accent. The result: age of first learning accounted for 59% of the variance in accent ratings, while how much they used English accounted for only 15%. And the decline with age was continuous — no cliff edge after which everything was lost. Strength of evidence: a large correlational study, not an experiment, so the number cannot be read straight off as causation. People who arrived early also differ in schooling, friendships, and work environments.

Why so stubborn? Two reasons, one on the movement side and one on the listening side.

The movement side is already on the table from "a few dozen milliseconds is the whole difference": the contrast is a sixty-millisecond relative timing, and you have never had direct control over that number. The instruction you can issue is "say this word", not "start the vocal folds fifty-eight milliseconds after the lips open". What needs changing sits several layers below the level at which you can give orders.

The listening side is worse. Flege's account is equivalence classification: if a sound in the new language is roughly like some sound in your first language, it gets filed into that existing cell, and so you never treat it as a new thing to be learned — you think you got it right, because you are listening to yourself with the old grid. That yields a counterintuitive prediction: the more similar, the harder. Sounds that are far from anything in your first language, and therefore sound plainly foreign, are more likely to be treated as new and given a cell of their own.

None of which means the system is frozen. Chang tracked adult native English speakers taking beginner Korean classes in 2012, and by the second week their voice onset times in English had already drifted toward Korean values, with more than one property shifting at once. Your native pronunciation keeps moving in adulthood; you just do not get to pick the direction. Strength of evidence: a small longitudinal sample, with the direction of the effect replicated in later work.

One last thing worth stating plainly. An accent is not a defect and not evidence of insufficient effort; it is the ordinary output of a system tuned early and afterwards nudged only by daily use. And the burden of poor communication does not sit entirely on the speaker's side. Clarke and Garrett found in 2004 that native English listeners hearing Spanish- or Chinese-accented English are indeed slower at first, but that the disadvantage narrows markedly within one minute, and in some conditions after only two to four utterances. Not understanding someone is a state that can improve in sixty seconds — not a fixed property of the person speaking.

🎒 In context · BigCat

① Working and living across languages: on a cross-border call you say a word, the other person misses it, so you say it again — louder, slower — and they still miss it. The section on the few dozen milliseconds explains why: they are cutting your sound with the boundary their own language drew, and repeating it just sends the same misfiled item through a second time. Volume and speed do not change which cell it lands in. What you can change is to stop repeating and pick one of three moves: swap in a synonym without that contrast, spell the word out, or shift the distinction from sound to vocabulary ("V, as in Victor"). The clue for choosing is what they echo back to you — the word they repeat tells you which cell your sound went into.
② Leading teams and organizations: in a review or hiring discussion someone says "he doesn't express himself clearly," and on further probing it turns out to mean the accent. "Thirty-two years in the country" supplies two facts you can pry apart with. Where an accent sits was settled by a very early time window and has no bearing on how clear the thinking is now; and adaptation on the listener's side is fast — the Clarke and Garrett result says one minute of exposure narrows the processing disadvantage markedly. Put together: in a fifteen-minute meeting, the first few minutes of effort get taken as a verdict on the whole thing. Two things are changeable. Break the rating items into things judgeable from a transcript — was the conclusion stated first, are there ambiguous sentences, was the question answered head-on — and move the material to before the meeting, so the adaptation happens before the evaluation starts.

🌀 Crossing over · Connections beyond the field

Distributed systems and AI · The criterion is relative timing, not an absolute moment: a speech recognizer trained on English will hear "pat" as "bat" from a Spanish-accented speaker. The mechanism is the one in the section on the few dozen milliseconds: the category boundary is drawn on the gap between two events, and where that line sits is each language's own choice. From this follows a conclusion you would not otherwise see: the direction of such errors can be worked out in advance, by laying the training language's boundary over the speaker's native distribution. It is not noise, it is a predictable systematic bias, which means part of it can be corrected by moving the decision threshold without retraining anything. The more testable half is the other one: a human listener's boundary shifts on its own, a frozen model's does not. Give thirty consecutive utterances from one speaker to a person and to a model — the person is more accurate on the thirtieth than on the first, the model gets the thirtieth wrong in exactly the way it got the first wrong.
Psychology · What is missing is the error signal, not the practice: motor skills usually improve with practice — free throws, cycling, touch typing. Accents can be practised for twenty years and hardly move. The difference is not effort but the equivalence classification from "thirty-two years in the country": the new target is filed into an old cell, so when you listen to yourself you cannot hear the deviation, and the repetitions never land where the change is needed. That gives a testable prediction: for equal training time, a group given immediate acoustic feedback — a screen showing where this attempt's voice onset time fell — should clearly beat a group that only repeats after a model, while simply doubling the practice volume should add almost nothing. Where this breaks down is the other direction: free throws are correctable by practice because whether the ball went in is an externally visible outcome, and that error signal never passes through your own perceptual system. An accent's error signal has to cross the very perceptual system that is the thing needing repair. So the rule "practice makes it better" holds only where the error criterion is independent of the perception being trained — once it is not independent, practice volume and improvement come apart.

Going deeper

If each language draws its own boundary, what does "correct pronunciation" mean?

Which cell a sound falls into depends on whose line is doing the cutting. So when someone is said to pronounce a word incorrectly, is the claim that they deviate from some community's median, or from an absolute reference? If only the former, the judgement travels with the community — the same person could move city and become "correct" without changing anything.

If similar sounds are harder, should teaching start by making them less similar?

Equivalence classification says a highly similar sound gets swallowed by an existing cell. Suppose early training deliberately exaggerated the difference, pushing the target somewhere plainly foreign — would learners find it easier to open a new cell? And if so, would step two, pulling it back into the normal range, just get it swallowed again?

Listeners adapt in a minute; why can't speakers adapt in decades?

One possible answer: listening means moving a classification boundary, speaking means rewriting a string of motor timings. But why should those two kinds of plasticity differ so enormously? Or is the listener's one-minute adaptation simply temporary, resetting with the next speaker, and therefore not the same sort of thing as the durable rebuild the speaker would have to do?