Electrophysiological evidence of statistical learning of long-distance dependencies in 8-month-old preterm and full-term infants

•We demonstrate 8month-old infants’ ability to compute long distance dependencies from an artificial speech stream.•We demonstrate that a frequency tagging approach can be used to explore online learning.•We show that temporal precision of syllabic neural processing seems crucial for statistical computations in speech. Using electroencephalography, we examined 8-month-old infants’ ability to discover a systematic dependency between the first and third syllables of successive words, concatenated into a monotonous speech stream, and to subsequently generalize this regularity to new items presented in isolation. Full-term and preterm infants, while exposed to the stream, displayed a significant entrainment (phase-locking) to the syllabic and word frequencies, demonstrating that they were sensitive to the word unit. The acquisition of the systematic dependency defining words was confirmed by the significantly different neural responses to rule-words and part-words subsequently presented during the test phase. Finally, we observed a correlation between syllabic entrainment during learning and the difference in phase coherence between the test conditions (rule-words vs part-words) suggesting that temporal processing of the syllable unit might be crucial in linguistic learning. No group difference was observed suggesting that non-adjacent statistical computations are already robust at 8months, even in preterm infants, and thus develop during the first year of life, earlier than expected from behavioral studies.