Is language learning a similar statistic ? In prosodic and phonological terms, babies learn the properties of their mother tongue extremely quickly. Even newborns just a few days old prefer to listen to sentences in their mother tongue - learning that relies mainly on slow, low-frequency prosodic information, and may begin in utero. Between six and twelve months, babies detect the phonetic categories of their language, and cease to distinguish contrasts that are irrelevant to them (e.g. the contrast between " r " and " l " for Japanese babies). This learning is clearly based on a statistical estimate of the distribution of the phonemes heard. Indeed, a bimodal distribution of phonetic features encourages the child to stabilize two distinct phoneme categories, whereas a unimodal distribution leads him to stabilize a single category. For vowels, learning would take place before six months, or even in utero. However, a recent experiment shows that this learning process is not accelerated in premature infants, even though they benefit from several extra months of exposure to language. There are therefore strong biological constraints on the maturation of cerebral language networks.
As far as the mental lexicon is concerned, the first evidence of recognition of spoken words is observed very early in the first year of life, well before the first words are produced. At five months, children prefer to listen to their own first name rather than that of another child, even if the latter shares the same tonal accentuation profile. Exposed for 10 days to 30 minutes of children's stories, children prefer to listen to this list of words they've heard before, rather than a list of new words.
