Word recognition reflects dimension-based statistical learning.

Abstract

Speech processing requires sensitivity to long-term regularities of the native language yet demands listeners to flexibly adapt to perturbations that arise from talker idiosyncrasies such as nonnative accent. The present experiments investigate whether listeners exhibit dimension-based statistical learning of correlations between acoustic dimensions defining perceptual space for a given speech segment. While engaged in a word recognition task guided by a perceptually unambiguous voice-onset time (VOT) acoustics to signal beer, pier, deer, or 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) to VOT. Results across four experiments are indicative of rapid, dimension-based statistical learning; reliance on the F0 dimension in word recognition was rapidly down-weighted in response to the perturbation of the correlation between F0 and VOT dimensions. However, listeners did not simply mirror the short-term input statistics. Instead, response patterns were consistent with a lingering influence of sensitivity to the long-term regularities of English. This suggests that the very acoustic dimensions defining perceptual space are not fixed and, rather, are dynamically and rapidly adjusted to the idiosyncrasies of local experience, such as might arise from nonnative-accent, dialect, or dysarthria. The current findings extend demonstrations of “object-based” statistical learning across speech segments to include incidental, online statistical learning of regularities residing within a speech segment.

Perception in a Variable but Structured World: The Case of Speech Perception

“…This provides qualitative evidence that listeners combine their prior expectations with observed cue distributions in order to rapidly adapt to unfamiliar talkers, as predicted by the ideal adapter framework . 2 More generally, unlike previous work showing constraints on distributional learning (Idemaru & Holt, 2011;Sumner, 2011), these results cannot be explained by a model where the only constraints on listeners' distributional learning is a binary distinction between natural and unnatural cue-category mappings. …”

“…Both studies involved gross, categorical mismatches between typical and experimental cue distributions: a reversal of the f0-to-voicing mapping in Idemaru and Holt (2011), and a remapping of /b/-like VOTs to /p/ in Sumner (2011).…”

The Role of Lexical Status and Individual Differences for Perceptual Learning in Younger and Older Adults

“…However, there are several studies that show sensitivity to distributional information that cannot be explained by selective adaptation (e.g., Clayards et al, 2008;Idemaru & Holt, 2011;Maye & Gerken, 2000), including in vowels (Liu & Holt, 2015). Thus it remains unclear whether distributional learning can reduce to selective adaptation and indeed, it has been argued that selective adaptation can be explained under the mechanisms of distributional learning (Kleinschmidt & Jaeger, 2015c) rather than the other way around.…”

Functional differences between statistical learning with and without explicit training

“…These results indicate that incremental adjustments to the language processing system occur on a continual basis, allowing for the dynamic and flexible acquisition of novel syntactic representations based on the current environment. Similarly, exposure to nonnative, accented speech causes learners to rapidly adjust their reliance on particular acoustic dimensions, a process referred to as dimension-based statistical learning (Idemaru and Holt 2011). This mechanism allows learners to adapt to the substantial acoustic variability among different speakers, accents, and dialects.…”