Early phonetic learning without phonetic categories: Insights from large-scale simulations on realistic input

Abstract

Before they even speak, infants become attuned to the sounds of the language(s) they hear, processing native phonetic contrasts more easily than nonnative ones. For example, between 6 to 8 mo and 10 to 12 mo, infants learning American English get better at distinguishing English and [l], as in “rock” vs. “lock,” relative to infants learning Japanese. Influential accounts of this early phonetic learning phenomenon initially proposed that infants group sounds into native vowel- and consonant-like phonetic categories—like and [l] in English—through a statistical clustering mechanism dubbed “distributional learning.” The feasibility of this mechanism for learning phonetic categories has been challenged, however. Here, we demonstrate that a distributional learning algorithm operating on naturalistic speech can predict early phonetic learning, as observed in Japanese and American English infants, suggesting that infants might learn through distributional learning after all. We further show, however, that, contrary to the original distributional learning proposal, our model learns units too brief and too fine-grained acoustically to correspond to phonetic categories. This challenges the influential idea that what infants learn are phonetic categories. More broadly, our work introduces a mechanism-driven approach to the study of early phonetic learning, together with a quantitative modeling framework that can handle realistic input. This allows accounts of early phonetic learning to be linked to concrete, systematic predictions regarding infants’ attunement.

Do Infants Really Learn Phonetic Categories?

“…This effect has been demonstrated both in cognitive models of auditory learning (Gauthier et al, 2007;Roark et al, 2020) and in machine learning models, where "pretraining" a system's perceptual space on a generic unsupervised task (such as predicting the next input in a sequence) can improve performance on a variety of downstream tasks (such as question answering or phone classification) (Chung et al, 2019;Devlin et al, 2019;Erhan et al, 2010;Peters et al, 2018;Schneider et al, 2019). While it is theoretically possible that systems pretrained on speech could be implicitly learning phonetic categories, evidence from models that do learn quantized representations (latent categories) suggests otherwise: the learned units are typically far more granular than phonetic categories, and often cannot even be well-characterized as sub-phones or subsets of phonetic categories Baevski et al, 2020;Chorowski et al, 2019;Hsu et al, 2021;Schatz et al, 2021).…”

“…4 Most models that have tested the feasibility of distributional learning for identifying phonetic categories have simplified the learning problem, for example, by using artificial data with low variability (McMurray et al, 2009;Pajak et al, 2013;Vallabha et al, 2007), focusing only on subsets of the categories infants would need to acquire (Adriaans & Swingley, 2017;de Boer & Kuhl, 2003;Gauthier et al, 2007), or limiting the training data to a single speaker (Miyazawa et al, 2010;Miyazawa et al, 2011). Similar models that were tested on more realistic datasets showed much worse performance at learning phonetic categories (Adriaans & Swingley, 2012;Jones et al, 2012;Schatz et al, 2021). Therefore, the distributional sensitivity that infants exhibit in simplified laboratory settings may not be sufficient to learn phonetic categories in naturalistic settings.…”

“…Methods for working with speech in naturalistic settings have been developed in the context of engineering applications, and naturalistic speech corpora now exist in numerous languages. By adapting these tools (e.g., Räsänen, 2011;Räsänen & Rasilo, 2015;Schatz, 2016;Schatz & Feldman, 2018;Schatz et al, 2013Schatz et al, , 2021, cognitive scientists can begin investigating the role of perceptual space learning in explaining how infants' perception of speech becomes specialized for their native language. 5 Thus far, we know of only a handful of models that have been evaluated against infant behavioral data after training on natural continuous speech.…”

“…5 Thus far, we know of only a handful of models that have been evaluated against infant behavioral data after training on natural continuous speech. Schatz et al (2021) trained a bottom-up distributional learner-specifically, a Dirichlet process Gaussian mixture modelon low-level spectral representations of speech from Japanese or English. The model reproduced infants' discrimination of [ r ] and [l], but the units it learned did not resemble phonetic categories.…”

“…phonetic category learning. Early phonetic category learning has been questioned before ( Jusczyk, 1992), yet only a few concrete alternative accounts of infants' changes in discrimination have been proposed (Guenther & Gjaja, 1996;Herrmann et al, 1995;Matusevych et al, 2020;Schatz et al, 2021).…”