Find an English speaker, have them hold a palm in front of their mouth, and say pin, then spin. The p in pin puffs air onto the hand; the p in spin barely does. Ask whether those two p sounds are the same and they will almost certainly say: same thing, both p.
But the difference on your palm is real, and it is exactly what separates two Mandarin words: 八 (bā, "eight") has the unpuffed one, 趴 (pā, "to lie face down") has the puffed one. The same pair of sounds is a boundary between two words in one language and simply doesn't exist in another.
It runs the other way too. 西 (xī, "west") and 诗 (shī, "poem") are two different words in Mandarin, while English has only one sh in that region — so English learners of Mandarin routinely file both into the same box. The ear receives the same thing; what gets heard differs. In between sits one step: sorting. What follows is what that grid looks like, why some people have no line between L and R, and when the lines get drawn.
Two terms first. A phoneme is the smallest unit of sound a language uses to tell meanings apart. The actual sounds that come out of a mouth, if swapping one for another never changes which word you said, are variants of a single phoneme — allophones. The test is crude: can changing just this one sound give you a different word?
Run the test. In Mandarin, swapping the initial of 八 (bā, "eight") for the initial of 趴 (pā, "to lie face down") changes the word — two phonemes. In English, putting the unaspirated p of spin where the aspirated p of pin belongs sounds odd, but nobody thinks you said a different word — one phoneme, two variants. The same pair of physical sounds, entirely different status in the two languages.
There is a very direct piece of evidence that the grid exists. In 1957 Liberman and colleagues synthesised a string of sounds spaced evenly along one axis: a clear b at one end, a clear p at the other, a dozen or so steps in between, each step exactly the same physical distance from the last. (The axis is the quantity from the earlier issue on the invisible choreography in your mouth — how many milliseconds separate the lip release from the onset of voicing.) Then they asked people to do two things.
The first was to listen one at a time and report b or p. If the ear reported the physics honestly, you would get a gentle ramp. What you get is a step: the first several stimuli are called b a hundred percent of the time, and after one particular step they flip to p a hundred percent of the time, with almost no transition zone.
The second task is the telling one: compare stimuli in pairs and only answer "same or different". Adjacent steps are equally far apart everywhere on the axis, so in principle every pair should be equally hard. In fact: two sounds on the same side of the step are nearly indistinguishable, with accuracy close to guessing; step across the boundary and suddenly the difference is obvious.
The result has been reproduced countless times; it is bedrock. But its strong version — that within-category differences are literally inaudible — was overturned by finer methods. In 2002 McMurray, Tanenhaus and Aslin put four pictures on a screen and had people click the word they heard, while nudging the word's initial consonant around inside a category. The word people reported never changed, but their eyes flicked toward the competing picture — and the closer to the boundary, the more they flicked, continuously.
So the more accurate statement is: what got flattened is your judgement, not the signal arriving in your head.
That Japanese speakers struggle with English r and l is famous enough to be a stereotype. It is worth telling because one experiment eliminated the "bad ears" explanation cleanly. First, where the two sounds differ: the shape of the vocal tract makes energy pile up at certain frequencies, and those bands are called formants. The main difference between r and l sits in the starting height of the third formant from the bottom — low for r, high for l — after which each glides into the following vowel.
In 1975 Miyawaki and colleagues synthesised a string running from ra to la, changing only that formant's starting point, evenly spaced. Japanese listeners doing the discrimination task produced a nearly flat curve; English listeners produced a clear peak at the boundary, the dashed line from the previous section.
The second step is the key one. They pulled that formant out on its own and stripped everything else away. What remains sounds like a chirp — nothing like speech. Same two groups, same physical differences, discrimination task again: the two curves nearly coincide. The Japanese group hears it just as well as the English group.
That settles it: the auditory system itself is not different. The difference appears at the step where the sound is handled as language — failing to hear it is not a failure to receive it, but a matter of what box it gets filed into once received. Evidence strength: a clean controlled experiment with synthesised speech, redone many times since.
And the Japanese side is not "missing a sound" either. Japanese has a single liquid phoneme, produced close to a tap — the tongue tip flicks once against the ridge behind the teeth, like the fleeting tt in American English butter. It sits between English r and l, so both English sounds fall into that one box. Korean likewise has a single liquid phoneme with two variants: an l-like sound in the syllable coda, as in 달 (dal, "moon"), turning into a tap between vowels, as in 다리 (dari, "leg" or "bridge").
The street runs both ways, so don't look down only one of them. English has one sh across the palatal-to-retroflex region; Mandarin draws two lines there — x as in 西 (xī, "west") and sh as in 诗 (shī, "poem") — and English speakers learning Mandarin have to add a line where they never had one. Every language has contrasts other people cannot hear. Nobody has the better ears.
Can a line be added later? Yes, but not cheaply. In 1991 Logan, Lively and Pisoni trained Japanese speakers to identify English r and l using materials from multiple talkers in multiple phonetic contexts; after a dozen-odd sessions, scores rose clearly, and the gains transferred to unfamiliar talkers and new words. Evidence strength comes in two layers: that training works at all is solid; but the specific recipe — more variety is better — failed to replicate in a 2022 study with 166 participants, whose authors concluded the advantage, if real, is much smaller than believed. Right direction, dosage undetermined.
When does the grid get drawn? Earlier than most people guess, and the process runs backwards: not learning new ones, but switching off the unused. In 1984 Werker and Tees tested infants in English-speaking homes. At six to eight months they could tell apart the two Hindi t sounds (one with the tongue tip against the back of the teeth, one with the tip curled up to the front of the hard palate), and also the two ejective stops of Nlaka'pamux, a Salish language of British Columbia (one made with the tongue body at the soft palate, one further back at the uvula). English-speaking adults can do neither. Then they tested ten-to-twelve-month-olds: also gone. In those few months, the net had closed.
But use the word "lost" carefully. Over that same span, discrimination of the contrasts their own language uses gets better. This is not decay, it is reallocation — the technical name is perceptual reorganization: a system that only has to handle English moves its resources onto the handful of lines English actually uses.
How is it learned? Nobody tells a baby which sounds count as one box, and nobody hands out labels. In 2002 Maye, Werker and Gerken played six- and eight-month-olds two minutes of sounds from one continuum, with a single difference between groups: the shape of the distribution. One group heard the endpoints often and the middle rarely (two peaks); the other heard the middle most (one peak). Afterwards only the two-peak group could tell the endpoints apart. Two minutes, no feedback, distribution shape alone — which incidentally explains the narrowing itself: a line your language does not use never produces two peaks in the input.
Run it in reverse: pour in a new language's distribution before the window closes, and a box should be addable. That is exactly what Kuhl, Tsao and Liu did in 2003. Nine-month-old American infants got twelve sessions of twenty-five minutes each, with live Mandarin speakers reading and playing with them. Afterwards they could tell apart the Mandarin contrast between q as in 七 (qī, "seven") and x as in 西 (xī, "west") — a pair English does not use. The same material delivered on a television screen, or as audio only, in the same room, for the same duration, on the same schedule, left infants indistinguishable from a control group that had never heard Mandarin at all. Evidence strength: a small-sample experiment whose direction has since been supported more broadly, though the conditions have been narrowed — a 2018 study found infants do pick up some of it from video when a peer is present. One controversy worth flagging: the narrowing itself is solid, but the perceptual magnet effect Kuhl proposed to explain it (the most typical sound of your language pulling nearby sounds toward it) is still argued over, and in 1998 Lotto and colleagues showed it is quite sensitive to task design.
One last thing to be clear about: the grid is not a prison. The training studies in the previous section show adults can still move it, and the eye-tracking result from the section on differences you flattened yourself shows the within-category information was never destroyed in the first place. The real constraint is how you move it: redrawing a line takes a changed input distribution plus external feedback, not listening harder — the "you" who is listening harder is the very layer that got flattened.
The report says same; the eye movements say different. So is "this person can't tell these two sounds apart" a statement about their perception, or about what they can report? Switch to a measure that doesn't rely on verbal report, and how many cases of "can't tell" turn into "can tell but can't say"?
If adults can partly get it back through training, what was switched off clearly isn't the receptor. Was the output of the sorting step wired shut, or was the channel to the reporting layer cut? The two stories predict different things — the second predicts that non-report measures should always show a residue.
At least three candidates: a live person tunes what they say to the child's attention; a live person supplies contingent feedback; a child is simply more engaged when someone is present. Can the three be pulled apart experimentally? Add real-time responsiveness to video — a live person on a video call — and where should the effect land?