Overheard speech and vocabulary

Why does speech aimed at a child build vocabulary when an equal amount of household talk overheard nearby does not?

Put a toddler in a room where adults talk to each other for an hour. Count the words: thousands of them, in full sentences, covering a wide vocabulary. Now put the same toddler with one adult who talks to her for an hour — fewer words, simpler ones, more repetition.

The second hour does far more for her vocabulary. That is awkward, because on any measure of quantity the first hour is richer. So the useful thing in speech cannot be the speech. What is it?

Go back to what a child hearing a new word actually has to solve. A sound arrives; the world contains a hundred things it might name. Nothing in the sound narrows that. The narrowing has to come from somewhere outside the word — from knowing what the speaker is talking about.

So ask: how does a child ever know that? Not by scanning the room, because the room is full of candidates. By watching the speaker. Where is she looking, what is she holding, and what is she doing it about? A word becomes learnable at the moment the child can locate the speaker's attention and find it pointed at something. That state — adult and child attending to the same thing, each aware the other is — is what makes reference solvable.

Now apply that to overheard talk. Two adults are talking to each other, and their topic is very often not in the room at all — yesterday, next week, a person elsewhere, an abstraction. The child can locate their attention perfectly well and find that it points at nothing she can see. The evidence needed to fix a word's meaning is simply absent, and no amount of additional overheard sentences supplies it, because they are all missing the same ingredient.

There is a second ingredient, perhaps the more important: contingency. When an adult talks to a child, what she says next depends on what the child just did. The child looks at the cup and hears "cup" within a second or two; she babbles and gets an answer. That timing is itself information — it tells the child which of her own acts the words respond to. Overheard talk has no such link; nothing the child does changes it.

Is contingency separable from content, or am I describing the same thing twice? It has been tested. Toddlers learn strikingly little from recorded video compared with a live person delivering the same material — the video deficit. But when the same content comes over a live video call, where the adult responds to that child in real time, learning improves relative to a recording. Same screen, same face, same words; only contingency differs.

That experimental form matters, because the correlational version of the claim is worthless on its own. Parents who direct more speech at their children differ from those who do not in education, income, working hours, stress, and everything those travel with — including whatever they pass on genetically. A raw association between directed talk at two and vocabulary at five is exactly what selection would produce with no causal arrow at all. The evidence that moves the question is the kind that manipulates contingency and watches learning move. Where only observational data exist, the honest answer is that the size of the effect is unknown, and I will not put a number on it.

What would refute the account? If a child learned a novel word as well when an adult named it while the child was attending to something else as when both were attending to the same object, the shared-attention leg collapses. And if live and recorded delivery of identical content produced equal learning under proper randomisation, contingency is not the mechanism.

Three refinements connect the steps.

First, "directed" is doing two jobs worth keeping apart. Directed speech tends to be about the here and now, which solves reference; and it is responsive to the child, which supplies timing. Both are missing from overheard talk, so ordinary observations cannot tell you which is carrying the weight. The video-call comparison isolates the second; isolating the first needs an experiment that holds contingency constant and moves only whether the referent is present.

Second, this is not the claim that ambient speech is useless. Children plainly pick up phonology, rhythm, and eventually vocabulary from talk not addressed to them, and in many communities a large share of a child's input is exactly that — children in such settings learn to talk perfectly well. What the evidence supports is narrower: for the youngest children, at the stage where reference must be solved from scratch, the directed subset predicts vocabulary growth where ambient volume does not. Recalled here: in one study of Spanish-speaking families, child-directed speech quantity predicted later vocabulary while overheard speech quantity did not. I am confident of the direction and not of any effect size, so no number is offered.

Third, this is a different claim from why an early vocabulary gap widens over the years. That argument is about growth depending on the stock you already have. This one is upstream of it. The fixed point of difference: there, identical exposure yields different amounts because of what the child brings; here, different kinds of exposure differ in whether they contain the evidence a word requires.

Finally, plainly: these are population averages across many families. They support no inference about any particular child or parent, and no instruction about how anyone ought to talk at home.

• What isolates presence-of-referent from responsiveness, since ordinary directed speech confounds them.
• How children in high-ambient settings compensate.

Every term the collection defines is gathered in the glossary.