The first 1000 days: Modeling language learning from individual children's everyday speech

Children learn to understand and speak from the speech around them, yet how that everyday input shapes learning remains unresolved. Large language models show that language is learnable from data, but their scale and training conditions differ sharply from a child's natural experience. We trained cognitively feasible, child-specific models on longitudinal home recordings from eight children across their first 1000 days of life. Each model learned incrementally from a single child's ordered speech input without prior linguistic knowledge, developing phoneme-like and word-form representations that tracked each child's learning pace. Learning was dependent on dense daily sampling and replay of recent input, indicating that both experience- and memory-related processes are necessary. These findings show that everyday experience contains sufficient structure for language learning and establish child-specific modeling for testing developmental theories.