Proceedings

Modeling Prosodic Development with Prenatal Audio Attenuation

---

Abstract

The delayed language development of preterm infants has been associated with diverse factors, from socioeconomic to neurological to linguistic. The current study explores the extent to which the duration of prenatal auditory exposure influences infants’ prosodic learning. Convolutional neural networks were trained on low-frequency speech audio followed by full-frequency speech audio to perform stress and tone classification, simulating prenatal and postnatal learning of prosodic patterns. We found that longer low-frequency exposure conferred an immediate advantage in stress and tone classification, although the advantage diminished in the long run. The simulation results highlight how sufficient low-frequency sound exposure impacts infants’ prosodic abilities. Surprisingly, training on full-frequency audio for the same duration as low-frequency audio gave rise to superior learning accuracy in our models, suggesting that acoustic cues beyond the low-frequency range can be informative for stress and tone categorization. Our findings emphasize the importance of considering both the quantity and quality of linguistic input in the infants’ early prosodic acquisition.

Keywords: Neural network modeling, Prosodic learning, Preterm infants, Convolutional neural network, Classification

How to Cite:

Tan, F., Zheng, S., Liu, M. & Do, Y., (2026) “Modeling Prosodic Development with Prenatal Audio Attenuation”, Proceedings of the Annual Meetings on Phonology 2(1). doi: https://doi.org/10.7275/amphonology.3665