Learning mechanisms in cue reweighting

Short-term perceptual re-weighting in suprasegmental categorization

“…Though there remain many open questions even for adaptive plasticity across segmental speech categorization, in general patterns of generalization seem to be consistent with an account that category activation drives reweighting (Idemaru & Holt, 2011; Zhang et al, 2021; Wu & Holt, under review; Wu, 2020) and the mechanism for this may involve supervised error-driven learning (Guediche et al, 2014; Wu, 2000) or reinforcement learning (Harmon et al, 2019). To the extent that short-term input regularities across acoustic dimensions are effective in activating a suprasegmental category even as they deviate from long-term expectations of correlations among input dimensions, we would anticipate re-weighting and modest generalization.…”

“…For instance, since English listeners perceptually weight VOT more than F0 in clear speech, VOT continues to signal /b/-/p/ category membership even in the context of a reversal of the VOTxF0 correlation in short-term speech input that conveys an accent. Several lines of evidence (Idemaru & Holt, 2011;Zhang et al, in press;Wu & Holt, under review;Wu, 2020) demonstrate that activation of an existing linguistic representation is crucial to eliciting the dynamic re-weighting of secondary acoustic input dimensions, perhaps through supervised error-driven learning (Guediche et al, 2014;Wu, 2000) or, alternatively, reinforcement learning mechanisms (Harmon et al, 2019). In sum, there is substantial evidence that segmental speech perception involves contributions from multiple acoustic dimensions, and that the relative perceptual weights of these contributions are dynamically adjusted according to listening context.…”

Changes to perceptual assimilation following training

“…As described in Kabakoff et al (2020), past studies incorporating passive distributional learning paradigms to represent how new speech sound categories are formed have evolved into a line of research attempting to optimize training paradigms for learning non-native contrasts. As such, the plentitude of studies demonstrating the effectiveness of feedback when learning new speech sound categories supports the addition of feedback to previous distributional training approaches (Goudbeek et al, 2008;Harmon et al, 2019;McCandliss et al, 2002).…”

Planting the seed for sound change: Evidence from real-time MRI of velum kinematics in German

“…It is this change in the proportionality by which nasalization is distributed between V and a following N in VNC ̥ that might then form the basis of a perceptual reweighting of cues. It is of course well known that there are multiple cues to speech sounds that have different perceptual weights (Clayards 2018, Francis et al 2000 and that listeners can reweight these cues in response to a change in the acoustic input (Francis & Nusbaum 2002, Harmon et al 2019. We suggest that the first stage of the sound change arises if the (proportional) increase of nasalization in the vowel due to the diminished N leads to a reweighting of perceptual cues such that listeners begin to pay greater attention to the nasalization in the vowel.…”

“…The starting point for Beddor's model is the well-known finding that there are multiple cues to speech-sound contrasts (Francis et al 2000, Harmon et al 2019, Holt & Lotto 2006, Lisker 1986) and that listeners vary in the attention or weight that they assign to these cues for disambiguating speech sounds (Beddor 2012, Chandrasekaran et al 2010, Clayards 2018, Kim & Clayards 2019, Schertz et al 2015. Compatibly with these findings, Beddor (2012Beddor ( , 2015 has shown that there are variable perceptual strategies for identifying nasalization in ṼN sequences (where Ṽ signifies a vowel with some coarticulatory nasalization): some listeners base their judgments mostly on information in N, others associate nasalization with ṼN (without parsing such judgments with either just Ṽ or just N), while yet others' perceptions of nasalization are swayed to a greater extent by information in Ṽ alone (see Stevens & Reubold 2014 for a similar argument for preaspiration).…”