Word extraction from continuous speech in newborns based on statistical information

Fló, A. , Ferry, A. & Mehler, J.

Cognitive Neuroscience Sector, SISSA, International School for Advanced Studies, Trieste Italy

In this experiment we ask if newborns can extract words from a continuous speech using statistical information. Previous work (Teinonen et al, 2009) demonstrated that newborns could detect word boundaries using just statistical information. Yet it is an open question if they actually extract the words or if they are only detecting boundaries. From previous studies using near-infrared spectroscopy (NIRS) we know that newborns show increased brain activity for a novel word than from a previously learned word (Benavides-Varela et al, 2011). Here, we used a similar methodology, asking if infants show different patterns of brain activity for words compared to part-words from continuous speech with only statistical cues. Infants were exposed to 3 minutes and 40 seconds of continuous speech. The speech stream consisted of four three-syllabic words of the form CVCVCV (e.g., ?lamipeduvokanubefi?). After familiarization, infants were presented with four test phases, each consisting of a short re-familiarization block followed by a test block, with silences of 25-30 seconds interposed between each block. In two of test blocks the four words from the stream were randomly played (e.g., ?lamipe?duvoka?), spaced by 0.5 - 1.5 seconds and repeated twice each, whereas in the other two, four part-words were presented (e.g., ...mipedu?.kanube...). If the infants extract and remember the words, we expect increased activity for the part-word blocks compared to the whole-word blocks. Initial results (N=27) show a differential activation in the test with words respect to the ones with part-words. Whereas part-words elicited an increase in oxyHb spread all over the temporal right hemisphere and in some fronto-temporal channels of left hemisphere, no increase was observed for words. These findings suggest that newborns can extract and remember four different words from continuous speech using only statistical information.