How Babies Learn Language: AI Models Trained on Eight Toddlers’ First 1,000 Days

How do babies go from babbling to hundreds of words in just a few years? It’s one of the oldest puzzles in science. A new peer-reviewed study in Science offers fresh clues about how babies learn language. Princeton researchers built a separate AI model for each of eight children, trained only on the speech that child actually heard at home, and found that each model learned at roughly the same pace as its real-life toddler.

It’s a modeling study, not a parenting guide, but it gives scientists a powerful new way to test ideas about early language.

Why this question is so hard

Children learn to understand and speak from the language around them, yet how that everyday experience shapes learning is still an open question.

Today’s large language models prove language can be learned from data, but they’re trained on far more text than any child ever hears, and in ways nothing like a child’s life. So they can’t tell us much about how real children learn.

How the study worked

Thousands of hours of real home audio

The team used recordings from the First 1000 Days project, which captured thousands of hours of everyday speech in the homes of eight children during their first 1,000 days of life.

One model per child

The researchers built a separate model for each child. Each model started with no prior knowledge of speech or words and learned only from the speech available to that child.

Learning on a child’s schedule

Training was designed to resemble a child’s life:

• Speech was processed day by day, in the order the child heard it.
• Daily exposure grew over time, as children spend more time awake as they get older.
• At the end of each day, the model replayed recent speech, to approximate the role sleep may play in memory.

What it shows about how babies learn language

The results were striking:

• Sounds first, then words. The models learned to tell English speech sounds apart within roughly the first simulated year, then went on to build vocabularies of thousands of words.
• Matching each child. Vocabulary growth in each model closely tracked that child’s measured vocabulary.
• One framework for all. The same approach worked across all eight children without modification. Three internal settings captured differences in how fast each child learned.

Two ingredients mattered most

1. Dense input. Models given only a fraction of a child’s daily recordings learned far fewer words, likely because many everyday words are rare and need repeated exposure.
2. Nightly replay. Models with little or no replay struggled to learn words, even when their daily input was rich.

The takeaway: the everyday speech around a single child, paired with learning and memory processes inspired by biology, carried enough structure to support core parts of language learning.

Important limitations

• It’s a model, not a child. The AI doesn’t think or learn like a toddler in every way. It tests whether a specific kind of learning process could explain the patterns seen in real children.
• Core aspects, not everything. The authors say everyday speech provided sufficient structure for core aspects of language learning, such as sounds and word forms, not that it explains all of language development.
• Eight children. The findings come from a small number of families with very detailed recordings. More children and more diverse settings will be needed.
• Not parenting advice. While dense everyday speech mattered for the models, this study doesn’t test specific advice for parents.

Why it matters

This work offers scientists a new tool: a way to turn theories about how children learn into computational models that can be checked directly against real children’s development. That could help researchers better understand why children learn at different rates.

The team has shared its code publicly. The home recordings themselves are kept private to protect the families.

The bottom line on how babies learn language

This Science study shows that child-specific AI models, trained on the real speech around eight toddlers and given nightly replay, learned sounds and then words at a pace that matched each child. It doesn’t solve how babies learn language, but it gives researchers an exciting new way to test the answers.

Read the study in Science: The first 1000 days: Modeling language learning from individual children’s everyday speech.