About this srcd poster session

---

Panel information
Panel 11. Language, Communication

Abstract

---

The influence of the prenatal acoustic environment on early phonetic learning (EPL)—infants’ specialization for the perception of the vowels and consonants of their native language(s)—is debated.

This influence may be substantial. First, a study reported that fetuses already specialize for the perception of native vowel contrasts (Moon et al., 2013). Furthermore, characteristics of the prenatal acoustic environment may explain otherwise puzzling features of EPL. Indeed, the mother’s body acts as a low-pass acoustic filter that better preserves cues to vowels than to consonants (Granier-Deferre et al., 2011a). This may explain the earlier specialization for vowels—around 6–9 months or earlier (Tsuji & Cristia, 2014; Moon et al. 2013)—than for consonants—around 10–12 months (e.g. Feldman et al. 2021). Finally, the low-pass speech sounds heard in utero have been proposed to provide infants with simpler input from which to learn, forming a natural curriculum, whose disruption—for example in preterm infants—may have detrimental effects on speech perception development (François et al., 2020; Vogelsang et al., 2023).

Several confounding factors and alternative explanations prevent definitive conclusions, however. The delayed specialization for consonant perception may reflect lower acoustic salience of cues to consonant contrasts or may result from maturational constraints (Menn et al., 2023), for example. Furthermore, delays in phonetic perception development in preterm infants are not systematically found (Gonzalez-Gomez et al., 2021) and could be explained by associated comorbidities, such as reduced exposure to language (Monson et al., 2023) or brain maturation delays (Phillips et al., 2011).

Computational modeling approaches may foster progress on this question. Modern machine learning technology enables explicit learning simulations at developmentally-relevant timescales. By controlling simulation parameters—such as whether speech input is initially low-pass-filtered—we can link them to their long-term observable consequences—such as whether specialization for the perception of consonants becomes delayed. This allows examining the feasibility of theoretical proposals and identifying decisive behavioral experiments.

Here, we perform large-scale learning simulations to understand the impact of initial exposure to the prenatal acoustic environment on an implementation of a distributional learning mechanism that reproduces hallmarks of EPL (Schatz et al., 2021). We also consider alternative implementations, to determine what is required for a learning mechanism to benefit from pretraining on low-pass filtered speech. To simulate prenatal experience, we pre-train models on appropriately low-pass-filtered speech (Dealessandri & Vivalda, 2018). We then train models on full-band speech to simulate postnatal experience. We instantiate models for different “native” languages and systematically vary prenatal and postnatal exposure duration. To measure trained models’ specialization to their “native” language, we use simulated phonetic discrimination tasks (Schatz et al., 2021). We hypothesize that pre-training with low-pass-filtered speech will result in (i) a delayed specialization for consonant perception relative to vowels and (ii) more accurate discrimination of native phonetic contrasts at convergence (curriculum effect).

This study constitutes a step towards a mechanistic understanding of the influence of the prenatal acoustic environment on EPL, contributing to our understanding of the computational mechanisms of perceptual attunement and informing the design of practical interventions on the acoustic environment of preterm infants.

Author information

Author	Role
Balleroy Ambre, Aix Marseille Univ, CNRS, LIS & CRPN, Marseille	Presenting author
Marxer Ricard, Université de Toulon & Aix Marseille Univ, CNRS, LIS, Marseille	Non-presenting author
Kabdebon Claire, Aix Marseille Univ, CNRS, CRPN, Marseille	Non-presenting author
Schatz Thomas, Aix Marseille Univ, CNRS, LIS, Marseille	Non-presenting author

⇦ Back to session