We are pleased to share a new publication from Frank, Shuang, Ming, and Youngah in the Proceedings of the Annual Meetings on Phonology. The paper, titled “Modeling Prosodic Development with Prenatal Audio Attenuation,” investigates how the sound environment before birth may shape early prosodic learning—the ability to perceive patterns such as stress and tone in speech.
Preterm infants often experience delayed language development, and one contributing factor may be the reduced duration of prenatal auditory exposure. To better understand this, the authors used convolutional neural networks to simulate infants’ early learning environment. The models were first trained on low‑frequency audio, reflecting the kinds of sounds fetuses can hear in utero, before being exposed to full‑frequency speech that resembles postnatal auditory input.
The study shows that longer exposure to low‑frequency audio provides an initial advantage for learning stress and tone patterns, though this early benefit fades over time. Interestingly, the simulations also reveal that learning improves even more when models are trained on full‑frequency audio for the same duration, suggesting that infants may rely on a wider range of acoustic cues than previously assumed. These findings underscore the importance of both the quantity and quality of auditory input in early prosodic development.
Thompson, A., Chik, A., Ngai, M., Chen, R., Ng, J., & Do, Y (2026). Iconicity and semantic transparency in Hong Kong Sign Language: Evidence from ratings and three guessing paradigms. Language and Cognition, 2(1). open_in_newDOI