Pre-Phonemic Layers Preserve Traces of Prior Languages in Speech Models

Speech models trained on a primary language before being switched to a second retain functional traces of the first in their lowest layers. Models with this early exposure re-learned the original language 14% faster than naive models, suggesting critical-period effects are driven by learning dynamics.