Neural correlates of statistical learning: discrimination of remembered and statistically possible items

Ordin, M. 1, 2 , Polyanskaya, L. 1 , Soto, D. 1, 2 & Molinaro, N. 1, 2

1 BCBL, Spain
2 IKERBASQUE - Basque Foundation for Science

Assume that the listeners are exposed to tri-syllabic sequences with high transitional probability so that syllable B follows syllable A and is followed by syllable C. So, the sequence ABC is segmented from continuous stream of concatenated syllables and is learnt as a whole constituent. Also, the acoustic stream contains syllables XY in some units, and YZ in different units. Both XY and YZ sequences are presented often, the sequence XYZ is never presented. Behavioral tests showed that XYZ nevertheless emerges as a perceptual unit, and is accepted as a word candidate (when you ask participant whether XYZ is a word from the language they heard, participants respond positively; and in a dual forced-choice test pitting random 3-syllabic sequences, ABC and XYZ sequences against each other). So, we have different types of statistical units, ABC-type is extracted from speech as whole constituents, XYZ-type are reconstructed based on the statistical regularities offline. Analysis of evoked potentials exhibited discriminable electrophysiological responses to the XYZ and ABC units and to random 3-syllabic concatenations.

Exploring the changes in neural oscillations during familiarization, we found that oscillatory changes at the frequency of discrete constituents reflect segmentation of continuous acoustic stream into separate units.