Integrating sensory input and prior expectations in structured auditory sequences

Abstract

Speech reaches the ear as a continuous stream of sound, from which listeners extract words without apparent effort. This ability rests substantially on statistical learning: on extracting the transitional probabilities between successive elements and using them to form expectations about the element to come. How the brain combines incoming acoustic evidence with such sequential expectations, and what that combination does to the neural representation of a predictable sound, is the subject of this dissertation. Two studies approach the question with different measures and different degrees of explicitness.

The first study shows that a learned association can be physiologically detectable without becoming apparent during learning or in behaviour. In two preregistered pupillometry experiments (n = 30, n = 31), spoken vowels were presented first in random order and then in fixed pairs. The pupil reliably followed the presentation of individual sounds but showed no reliable signature of the pair structure in any of the measures obtained. In a final recognition task, participants' judgements remained at chance, while pupil diameter reliably distinguished previously heard from newly recombined pairs — independently of what participants themselves reported.

The second study separates two effects that are hard to tell apart in two-element sequences. Thirty-six participants explicitly learned six pseudowords, each made up of three consonant–vowel syllables, which were then presented in continuous sequences during EEG recording, alongside sequences without word structure. A subset of syllables was an acoustically ambiguous blend of two learned alternatives. For these, the proportion of expectation-consistent categorizations rose linearly across the three positions – the influence of expectation increased as the word unfolded. Identification of clear syllables, by contrast, followed the contour of the word. It was better at the two word boundaries relative to the unstructured sequence, but not at the medial position – which is precisely where the neural expectation effects were strongest: amplitudes were reduced and syllable identity was already detectable before stimulus onset. At the word-final position the direction of the amplitude effect reversed, consistent with completion of the unit. Position-dependent predictability is therefore one of several factors shaping word processing in a sequence context.

Both studies show a similar pattern: in the first study, learned knowledge and its expression in behaviour diverge; in the second, neural and behavioural expectation effects are distributed differently across positions in word processing. This separation becomes clearly visible only in units of more than two elements. For research on statistical learning this means that the structure of the sequence co-determines which effects can appear at all, and that a single behavioural measure says little about what has been learned.