Listen to a conversation in a language you have never studied, and something unsettling becomes obvious: there are no spaces. Speech arrives as an unbroken acoustic stream, and its pauses fall as often inside words as between them. The neat gaps on this page are an invention of writing, not a property of sound.
Yet by around eight months of age, infants pull individual words out of this stream, before they know what most of those words mean. With no dictionary and no instruction, they solve what linguists call the segmentation problem, a task that has long challenged engineers building speech recognition systems.
How do they do it? Decades of psycholinguistic research point to one answer: infants are quietly tracking statistics. They register which sounds tend to follow which, which syllables carry stress, and which mouths move in time with the voice. Word boundaries, it turns out, are not handed to the learner. They are inferred from patterns, and that inference begins remarkably early in life.
Transitional Probability: Listening for Where Prediction Breaks
In 1996, Jenny Saffran, Richard Aslin and Elissa Newport ran an experiment that reshaped the field. They played eight-month-old infants two minutes of synthesised speech. It was a monotone stream of syllables like bidakupadotigolabu, with no pauses, no stress and no meaning. Hidden inside it were four three-syllable 'words' repeated in random order.
The only clue to where these words began and ended was statistical. Within a word, one syllable always predicted the next: bi was followed by da every time. Across word boundaries, prediction weakened, because any of the other words could come next. This measure, the likelihood that one syllable follows another, is called transitional probability. In testing, infants listened differently to the embedded words than to sequences that spanned a boundary. They had tracked the difference.
Natural language shows the same signature. In English, pre is followed by ty with high probability, because pretty is a common word. But ty is followed by ba only rarely. It happens in pretty baby and in few other places. A dip in predictability is therefore a reasonable guess at a boundary. Later studies found the same learning for musical tones, visual shapes, and even cotton-top tamarins. This suggests a general-purpose mechanism older than language itself.
The finding sharpened an old debate. Generative linguists following Chomsky had argued that children's input is too impoverished to support learning from experience alone. Statistical learning showed the input is richer than many had assumed, but it also has limits. Computational models by Charles Yang found that raw transitional probabilities perform poorly on real child-directed speech, where most words are a single syllable. Statistics, it seems, need structure to work with.
TakeawayInfants don't need to be told where words are; they notice where predictability breaks down. A word boundary, in this view, is simply the point where the future becomes uncertain.
Prosodic Cues: The Rhythm That Marks the Edges
Statistics are not the only signal in speech. Languages also carry prosody, the rhythm, stress and pitch contours layered over individual sounds. Prosody is the first part of speech infants hear clearly. In the womb, the mother's body filters out much of the fine detail of speech, but the melody passes through. Newborns can already tell apart languages that differ in rhythm, such as English and Japanese. They cannot yet separate two languages from the same rhythmic class.
In English, rhythm is a remarkably useful guide. Anne Cutler and David Carter found that about ninety percent of content words in conversational English begin with a stressed syllable: DOCtor, BASket, WAter. A listener who assumes every strong syllable starts a new word will be right most of the time. Cutler and Dennis Norris called this the metrical segmentation strategy.
Peter Jusczyk and colleagues showed that infants adopt this heuristic early. At seven and a half months, English-learning babies readily picked out trochaic words like kingdom from fluent passages. Iambic words like guitar gave them trouble. They treated the stressed tar as a word in its own right, often fusing it with whatever came next. By ten and a half months, with more cues at their disposal, they had corrected course.
The crucial point is that the strategy is specific to each language. French places prominence at the ends of phrases rather than at the starts of words, and French-learning infants segment accordingly. Erik Thiessen and Saffran found that seven-month-olds favour statistical cues when the two conflict, while nine-month-olds favour stress. One plausible reading is that babies use statistics to find their first words, then extract a stress pattern from those words. That pattern then speeds up everything that follows.
TakeawayLearning compounds: an early, imperfect method can uncover the examples that teach a better one. Infants appear to bootstrap their way from counting syllables to hearing rhythm.
Cross-Modal Integration: Words in a Social World
Laboratory studies often isolate speech from everything else, but real infants rarely hear voices in a vacuum. They see faces, watch hands, follow gazes, and feel the rhythms of being held. A growing body of research suggests that segmentation draws on all of these channels at once. The statistical machinery appears to work best when it is embedded in a social world.
Consider the face. George Hollich, Rochelle Newman and Peter Jusczyk found that seven-and-a-half-month-olds struggled to segment words from a voice masked by background talk. They succeeded when they could see a synchronised video of the speaker's face. Lip movements line up in time with syllables, giving the infant a second, visual record of where acoustic events begin and end. Even an abstract display pulsing in time with the voice helped, which suggests that synchrony itself carries information.
Familiar words act as anchors too. Heather Bortfeld and colleagues showed that six-month-olds could segment a new word when it followed their own name or the word Mommy. They could not do so when it followed an unfamiliar name. Each known word marks the edge of its neighbour, so vocabulary grows outward from a few stable footholds. Infant-directed speech, with its exaggerated pitch and slower pace, sharpens the statistical signal further.
Social context may be more than helpful; for some learning, it may be necessary. Patricia Kuhl found that American nine-month-olds exposed to Mandarin through live interaction kept their sensitivity to Mandarin sound contrasts. Infants who watched the same speakers on video did not. That study concerned speech sounds rather than word boundaries, but it points to a broader principle. The infant's statistical learner appears to be switched on, and tuned, by another person's attention.
TakeawayInfants do not learn from a signal; they learn from a situation. Every channel, whether sound, sight, touch or shared attention, narrows the space of possible interpretations.
The segmentation problem looks impossible only if we picture the infant as a passive receiver waiting for clean input. In fact, babies are active pattern detectors, weighing probabilities, rhythms and visual synchrony against one another.
No single cue is reliable on its own. Transitional probabilities blur in real speech, stress patterns mislead with iambic words, and faces are not always visible. Robustness comes from integration, the same principle that makes adult perception so resilient.
This research also reframes the old nature-versus-nurture question. The capacity to compute statistics may be innate and evolutionarily ancient, but what it computes depends on the language a child hears. Universal machinery with local tuning may be the deepest pattern of all.