Say a sentence out loud and it feels like a row of neat, separate words. It isn't. In a recording the gaps between words mostly vanish: the sound is a continuous smear, and the quiet moments you can see on a waveform often fall inside a word rather than between two words. A baby has to cut this stream into words before knowing a single word. How?
Here is a clue that is hiding in the sound itself. Inside a word, one syllable reliably follows another. After ba in baby, you will hear by nearly every time. But after by the next syllable could be almost anything: is, likes, for. Syllable pairs that are glued together are probably inside a word. Pairs that are unpredictable are probably a boundary.
The number for this is the transitional probability: out of all the times you hear syllable A, how often is it followed by syllable B? A value of 1 means "always". A low value means "A could be followed by lots of things".
In 1996 Jenny Saffran, Richard Aslin and Elissa Newport tested this on 8-month-old infants. They made four made-up words of three syllables each, such as bidaku, padoti, golabu and tupiro, and played them in a random order for about two minutes in a flat, monotone voice with no pauses, no stress and no hint of where words start. Inside a word the transitional probability was 1.0 (da always follows bi). Across a word boundary it was only about one in three, because any of the other three words could come next.
Afterwards the babies heard single items, either real words from the stream or "part-words" made of syllables that had straddled a boundary (the end of one word plus the start of the next). Infants turned their heads longer toward the unfamiliar-sounding items, which in this test were the part-words. That difference in attention shows the babies had registered which syllable sequences were solid and which were accidents of the order. Two minutes was enough.
Below is a stream built the same way: 20 words from the four above, in random order, printed as one unbroken run of syllables. The bar under each syllable is the learned transitional probability from the syllable before it. Use the first slider to choose how much of the stream you have listened to; only that part is used to learn. Then use the second slider to decide how low a probability must be before you cut the stream there.
Dimmed syllables are the ones you have not heard yet. Bars there still use what you learned from the earlier part, the way a baby uses past listening on new speech. Grey bars mean you never heard that first syllable.
Start with very little listening. Early on, a single odd coincidence looks like a rule, so you will cut inside words or miss boundaries. Drag the first slider up and watch the bars inside words climb to full height while the bars at boundaries settle near one third. Then try the cut-off: too low and you miss boundaries, too high and you start chopping words in half. Somewhere in the middle there is a range that works well, because within-word pairs stay at 1 while boundary pairs hover well below it.
Real speech is messier than four tidy words. It has stress patterns, pauses at phrase edges and sound rules that differ across languages, and infants use those cues too. The statistical trick is one tool among several, not the whole story. It is also not unique to language or to humans: learners, including adults, can pick up regularities in tone sequences and in pictures, and the effect has been found in other animals. What the experiment did show is that long before a child says a first word, the child's brain is already counting which sounds follow which, and using the counts to find the words it is about to learn.