Infant Phonetic Learning as Perceptual Space Learning: A Crosslinguistic Evaluation of Computational Models

Abstract

In the first year of life, infants' speech perception becomes attuned to the sounds of their native language. This process of early phonetic learning has traditionally been framed as phonetic category acquisition. However, recent studies have hypothesized that the attunement may instead reflect a perceptual space learning process that does not involve categories. In this article, we explore the idea of perceptual space learning by implementing five different perceptual space learning models and testing them on three phonetic contrasts that have been tested in the infant speech perception literature. We reproduce and extend previous results showing that a perceptual space learning model that uses only distributional information about the acoustics of short time slices of speech can account for at least some crosslinguistic differences in infant perception. Moreover, we find that a second perceptual space learning model, which benefits from word‐level guidance. performs equally well in capturing crosslinguistic differences in infant speech perception. These results provide support for the general idea of perceptual space learning as a theory of early phonetic learning but suggest that more fine‐grained data are needed to distinguish between different formal accounts. Finally, we provide testable empirical predictions of the two most promising models and show that these are not identical, making it possible to independently evaluate each model in experiments with infants in future research.

Modeling the initial state of early phonetic learning in infants

“…While this is expected in the native condition (e.g., exposure to Japanese makes the model better at discriminating Japanese sounds), this might be more surprising in the non-native condition. This is because there are many shared sounds across the two languages and the results reported in panel b) are computed across all possible contrasts -similar to what has been observed by Lavechin et al (2022), Matusevych et al (2023), andSchatz et al (2021).…”

“…By 10-12 months, American English infants show an improvement (facilitation) in their ability to discriminate the [ô]-[l] contrast, while Japanese infants show a decline (loss). While the effect of language exposure (higher scores for the model for whom the contrast is native) has been reproduced in numerous computational modeling studies and across different pairs of languages -e.g., Lavechin et al (2022), Li et al (2020), Matusevych et al (2023), andSchatz et al (2021) -, a closer examination of the trajectories taken by the proposed algorithms reveals notable differences with the trajectories observed in infants. Schatz et al (2021) used an algorithm based on a mixture of Gaussians applied to melfrequency cepstral coefficients (MFCCs) with their first-and second-order derivatives.…”

“…To demonstrate the relevance of our approach in modeling early phonetic learning, we simulate the learning process of American English-and Japanese-learning infants using CPC, an algorithm that learns from raw speech in an unsupervised manner already proposed in Lavechin et al (2022Lavechin et al ( , 2024 and Nguyen et al (2020) -see Matusevych et al (2023) for a comparison of different models. To induce 'innate' speech sound discrimination capabilities and propose models more aligned with the attunement or universal theories, we pretrain models on ambient sounds in Experiment 1, and on multilingual speech in Experiment 2.…”

“…We consider two types of models. One model starts from random initialization, which is akin to assuming little initial discrimination capabilities except those brought by the architecture which has been optimized to process human speech (see Rivière et al, 2020) and corresponds to how computational models of early phonetic learning are typically trained (e.g., see Lavechin et al, 2024;Matusevych et al, 2023;Schatz et al, 2021). This is our no-pretraining condition, which aligns with the perceptual learning theory.…”

Computational Modeling of the Segmentation of Sentence Stimuli From an Infant Word‐Finding Study

“…Our work here is, fundamentally, a demonstration project, joining other work that aims to harness developments in computational speech technology (e.g., Elsner, Goldwater, & Eisenstein, 2012;Matusevych, Schatz, Kamper, Feldman, & Goldwater, 2023;Räsänen & Rasilo, 2015;Roy & Pentland, 2002) with precursors in the machine learning literature (e.g., Harwath, Torralba, & Glass, 2016). We view our approach as contrasting with the direction that language acquisition research usually takes.…”