Eight-month-olds find the words in two minutes

In 1996, three researchers played eight-month-old babies two minutes of a language nobody speaks. It was a synthesized stream of four three-syllable nonsense words, tupiro and golabu among them, run together in random order with no pauses and no stress.

The only clue to where one word ended and the next began was statistical: within a word, each syllable always followed the last; across a boundary, it followed only a third of the time. The experiment is Saffran, Aslin and Newport (1996).

Then they tested. The babies heard the "words" again, and also new sequences that spanned a boundary, and they listened longer to the new ones.

They had noticed. After two minutes, with no instruction and no meaning, they had found the words.

This is statistical learning: picking up regularities in what you hear without being told there is anything to pick up. Later work found adults do it too, and that it works on tones as well as syllables. It is one of the basic machines a brain runs on speech, and it starts before a first birthday.

I want to keep this in proportion. Finding where words begin is one small skill, the made-up language made the cue unusually clean, and knowing where a word ends is not knowing what it means. The experiment does not say that hearing a language teaches you it.

What it says is narrower and, to me, more encouraging: the pattern-finding runs on its own, and what it needs is input. A language you do not yet understand sounds like one long stream because you have not heard enough of it for the machine to work.

Give it the hours. Aloglot is built to make those hours bearable: it reads you a story in the language while lighting each word as it is spoken, so the stream comes pre-cut, with the English under every sentence and any word explained on a tap while your ear catches up.

Sources

• Saffran, Aslin and Newport (1996). Statistical learning by 8-month-old infants. Science, 274(5294), 1926–1928.