Specificity of dimension-based statistical learning in word recognition.

Speech perception flexibly adapts to short-term regularities of ambient speech input. Recent research demonstrates that the function of an acoustic dimension for speech categorization at a given time is relative to its relationship to the evolving distribution of dimensional regularity across time, and not simply to a fixed value along the dimension. Two experiments examine the nature of this dimension-based statistical learning in online word recognition, testing generalization of learning across phonetic categories. While engaged in a word recognition task guided by perceptually unambiguous voice-onset time (VOT) acoustics signaling stop voicing in either bilabial rhymes, beer and pier, or alveolar rhymes, deer and tear, listeners were exposed incidentally to an artificial "accent" deviating from English norms in its correlation of the pitch onset of the following vowel (F0) with VOT (Experiment 1). Exposure to the change in the correlation of F0 with VOT led listeners to down-weight reliance on F0 in voicing categorization, indicating dimension-based statistical learning. This learning was observed only for the "accented" contrast varying in its F0/VOT relationship during exposure; learning did not generalize to the other place of articulation. Another group of listeners experienced competing F0/VOT correlations across place of articulation such that the global correlation for voicing was stable, but locally correlations across voicing pairs were opposing (e.g., "accented" beer and pier, "canonical" deer and tear, Experiment 2). Listeners showed dimension-based learning only for the accented pair, not the canonical pair, indicating that they are able to track separate acoustic statistics across place of articulation, that is, for /b-p/ and /d-t/. This suggests that dimension-based learning does not operate obligatorily at the phonological level of stop voicing.