A language you don't know always sounds impossibly fast, one unbreakable ribbon of sound, and you want to ask the speaker to slow down. But in your own language you speak no slower than they do — the listener simply has something you lack.
Three things this time: how a baby can cut words out of a seamless stream of sound, why counting alone isn't enough, and how far the chunk you cut out still is from being a word. The starting point is an unremarkable number: two minutes.
In 1996 Saffran, Aslin and Newport invented a language with four words in it: bidaku, padoti, golabu, tupiro. Three syllables each, all made up. A synthesiser then strung those four words together at random into one continuous run of syllables — no pauses, no stress, no rise or fall, delivered as flatly as a machine reading a list. For two full minutes an eight-month-old sat and listened.
The only clue in that stream is probability. Inside a word, bi is always followed by da and da always by ku: hear the first, and the chance the next one is that one is a hundred percent. At the seam between two words, whatever comes after ku is the start of any of the four, chosen at random — a chance of one in three.
After two minutes the testing begins. The method is the head-turn preference procedure: whichever speaker the infant keeps looking toward goes on playing; when she looks away, the sound stops. So "how long she listened" becomes a number you can measure. One speaker plays whole words like bidaku, the other plays three-syllable chunks cut across a seam, like kupado. She has heard all of those syllables; she has just never heard the second kind as one unbroken piece. The two listening times come out different. The difference is itself the answer: she has told the two kinds apart.
There is an obvious objection: perhaps combinations like bidaku simply came up more often and grew familiar. In 1998 the same group equated the frequencies and ran it again — both kinds of chunk occurred equally often, only the conditional probability differed, and the infants still told them apart. So what she computes is not "how many times have I met this" but "having met this, is the next one that".
The ability is not a human monopoly, and not specific to speech. Play the same synthetic streams to cotton-top tamarins and they tell the chunks apart too; it was later shown in rats. Infants do the same thing with sequences of pure tones. It looks less like a language faculty and more like a general-purpose sequence-statistics machine that language happens to use.
The evidence strength should be stated plainly. The original experiment has been replicated many times, but a 2017 meta-analysis paints a less flattering picture: among close replications the effect is significant but small; pool all studies that use conditional probability for segmentation and the overall effect is not significant; and synthetic versus naturally produced speech makes a visible difference. The phenomenon is real; the magnitude is less imposing than the textbooks suggest.
Real speech is not that machine-flat stream. It has strong and weak, long and short, high and low — and those can serve as signposts. Different languages hand out different signposts.
English hands out stress. Cutler and Carter, in 1987, went through two electronic dictionaries and a large body of real conversation and came out with a very useful proportion: in everyday conversation about ninety percent of content words — nouns, verbs, adjectives, not the little grammatical words — begin with a strong syllable. That makes one crude rule a good bet: when you hear a stress, assume a new word starts here.
How do we know English-learning infants really use that rule? Watch it fail. In English, guitar is one of the minority of weak-strong words, with the stress at the end. Jusczyk and colleagues found in 1999 that seven-and-a-half-month-olds hearing sentences like "guiTAR is…" pull out TARis — treating tar as the start of a word and sweeping the unstressed is in after it, assembling something that is not a word at all. A rule misapplied is a rule in use.
And that signpost belongs to English; it does not generalise. French words carry nothing like English's fixed-place stress, and French listeners segment by the syllable. Japanese segments by a unit smaller than the syllable, the mora — a sub-syllabic beat, so サッポロ (Sapporo, the city) is heard by Japanese listeners as sa-p-po-ro, four beats, the little doubled consonant taking a beat of its own. Mandarin is different again: nearly every syllable carries a tone and is pronounced in full, so the alternation of strong and weak simply does not arise. One job, four languages, four ways of measuring.
Earlier evidence still comes from newborns days old. In 1998 Nazzi and colleagues played French newborns low-pass filtered sentences — the crisp consonants and vowels smeared away, leaving only the outline of rhythm and intonation. The newborns could tell English from Japanese, but not English from Dutch. The first pair belongs to different rhythm classes, the second to the same one. Before any native language is in place, the infant is already gripping that coarse outline. One caveat: recent meta-analytic and modelling work questions whether rhythm is really the operative variable, so the mechanism is not settled.
The signposts are necessary because pure statistics is brittle. Johnson and Tyler changed exactly one thing in 2010: they made the words in the artificial language unequal in length, two two-syllable words plus two three-syllable words, with conditional probability still a perfect cue. This time the eight-month-olds failed. Once word length varies, counting neighbouring syllables alone cannot hold — and words in real languages of course vary in length, so the cues have to be used together.
Even if every cut is right, all the child holds is a pile of recurring sound chunks. To be a word, a chunk has to be the name of something. Does that step come free?
Graf Estes and colleagues joined the two halves together in 2007. Seventeen-month-olds first heard a stream to be segmented statistically, and were then immediately taught names for objects. The names came in three kinds: whole words from the stream they had just heard, chunks cut across a seam, and entirely new syllable sequences. Only the first kind was learned. The second kind is the telling one — it had occurred in the stream just as often, it had simply never stood as one whole piece.
So segmentation is not a by-product of word learning but the step before it: only a chunk that has been cut out is eligible to be a candidate name. That also explains something otherwise odd — why it is still worth listening when you understand nothing at all. What is being accumulated is a candidate list.
Does any of this hold outside the laboratory's tidy artificial languages? In 2009 Pelucchi and colleagues played eight-month-olds learning English, with no prior exposure to Italian, naturally produced Italian speech — a real voice, real intonation, real words. They still separated the higher- from the lower-probability syllable combinations. Statistical segmentation had walked out of the artificial stream.
But do not inflate the step either. Cotton-top tamarins can segment, and nobody has reported a tamarin going on to treat the chunk as the name of anything. So what actually needs explaining is not the statistics machine — that machine is widely distributed — but what comes immediately after it: why only humans readily take a stretch of sound to be the name of a thing. There is no agreed account of that step yet.
Adults agree fairly well about what counts as one word, yet the criterion is hard to state — is "go to work" one word or three? Statistical segmentation gives you not boundaries but an undulating curve, and the boundary is a line you draw on it. So: is there a natural place on that curve for the line, or is it set by the writing system, by printing convention, by how a teacher taught you in the year you learned to read? Japanese and Thai writing put no spaces between words at all.
There are at least three candidate answers: the chunks do not persist in memory for them, the chunks cannot be made to stand for arbitrary things, or chunks cannot be combined into larger structures. Each defect would leave a different behavioural trace. What task would you design so that the three explanations make different predictions, rather than vaguely testing "can it learn words"?
One reading is that the phenomenon itself has been overstated. Another is that the studies differ too much (synthetic or natural speech, length of exposure, age of infant) for the average to mean much. The two readings call for different next steps: the first for a preregistered large-sample direct replication, the second for stratifying by variable before looking again. Given a pooled result with high heterogeneity, how do you decide which road to take?