Simultaneous tracking of coevolving distributional regularities in speech.

Abstract

Speech processing depends upon mapping variable acoustic speech input in a manner that reflects the long-term regularities of the native language. Yet, these mappings are flexible such that introduction of short-term distributional regularities in speech input, like those arising from foreign accents or talker idiosyncrasies, leads to rapid adjustments in the effectiveness of acoustic dimensions in signaling phonetic categories. The present experiments investigate whether the system is able to track simultaneous short-term distributional statistics present in speech input, or if instead the global regularity jointly defined by these distributions dominates. Three experiments establish that adult listeners are able to track distinct simultaneously-evolving regularities across time, given information to support the ‘binning’ of acoustic instances. Both voice quality and visual information to indicate talker supported tracking of co-evolving distributional regularities, even when the regularities were opposing and even when the acoustic speech tokens contributing to the distinct distributions were identical. This indicates that re-weighting of perceptual dimensions in response to short-term regularities in speech input is not simply an accumulation of acoustic instances. Rather, the system is able to track multiple context-sensitive regularities simultaneously, with rapid context-dependent adaptive adjustments in how acoustic speech input maps to phonetic categories.

Generalization of dimension-based statistical learning

“…Prior work thus intriguingly suggests highly specific learning arising across distributions of acoustic dimensions that is resistant to phoneme generalization and supports tracking of independent regularities across speech categories, even for the same talker (Idemaru & Holt, 2014;Maye & Gerken, 2001;Zhang & Holt, 2018). The current study further examines generalization in dimension-based statistical learning.…”

“…Our prior work has repeatedly shown that the perceptual weight of the F0 dimension in voicing categorization decreases in Reverse block (Idemaru & Holt, 2011Zhang & Holt, 2018). But, we have always tested perceptual adjustment of F0 using VOT-neutral test stimuli (e.g., VOTneutral beer/pier) that aligned in the word frame with exposure stimuli (accented [b] and [p] in beer and pier).…”

“…Listeners continued to rely on F0 in voicing categorization at another place of articulation (e.g., alveolar [d] vs. [t]) even though all stimuli were produced by the same talker and the sounds occurred in the same phonetic context (i.e., [beer]/[pier] and [deer]/[tear]). Furthermore, when listeners experienced competing F0/VOT statistics across two places of articulation within the same block of trials, they showed evidence of tracking independent statistics at each place of articulation (see also Zhang & Holt, 2018). In experiencing a F0/VOT correlation reversal for [b] and [p] along with the canonical F0/VOT correlation for [d] and [t], the cumulative statistics across place of articulation nullify the F0/VOT correlation at each place such that there is no F0/VOT correlation in the short-term input.…”

“…In encountering short-term speech input that deviates from longterm regularities, such as in a foreign accent, perceptual weights rapidly adjust. Previous research (Idemaru & Holt, 2011Lehet & Holt, 2017;Schertz, Cho, Lotto, & Warner, 2016;Zhang & Holt, 2018) has shown that when the long-term English relationship of F0 to voicing categories, as expressed in beer versus pier and also deer versus tear, reverses in local, short-term speech input creating an artificial "accent," listeners rapidly down-weight reliance on F0 in speech categorization. This dimension-based statistical learning has been demonstrated across spectral and durational acoustic dimensions signalling vowel categories as well (Liu & Holt, 2015) and influence listeners' own productions (Lehet & Holt, 2017).…”

“…The previous work on dimension-based statistical learning contributes to a growing literature that has demonstrated that speech categorization is adjusted dynamically in response to a short-term deviation from the norm temporarily experienced in the speech signal. We now know that different types of information can guide such perceptual adjustment, including lexical information (Eisner & McQueen 2005;Kraljic & Samuel, 2006Norris, McQueen, & Cutler, 2003;, visual information (Bertelson, Vroomen, & de Gelder 2003;Reinisch, Wozny, Mitterer, & Holt, 2014;Reinisch & Holger, 2016;Vroomen, van Linden, de Gelder, & Bertelson, 2007;), phonotactic information (Culter, McQueen, Betterfield, & Norris, 2008), and statistical distributional information from the acoustics (Clayards et al, 2008;Idemaru & Holt, 2011Liu & Holt, 2015;Zhang & Holt, 2018).…”