How do babies learn language so quickly? A new AI study may help scientists answer this difficult question.

Researchers at Princeton University recorded the daily lives of children from birth to about three years old—their first 1,000 days. Cameras and microphones in their homes captured conversations, family activities, and other everyday sounds. The researchers then created a separate AI model for each of eight children. Each model received the speech heard by one child, in the same order the child had experienced it. (pni.princeton.edu)

At the beginning, the AI had no dictionary and no knowledge of English sounds. It was not even told where one word ended and another began. During its first simulated year, however, it started to recognize different speech sounds. Later, it found repeated word forms and built a vocabulary of thousands of words. The speed of its vocabulary growth was similar to that of the child whose recordings it used. (hassonlab.princeton.edu)

One important discovery involved memory. Simply listening to new speech every day was not enough. At the end of each simulated day, the AI needed to “replay” some recent speech. This was designed to copy one possible role of sleep: helping the brain strengthen memories. When the AI had little or no replay, it struggled to learn words. It also learned less when it received only a small part of the child’s daily speech. (pni.princeton.edu)

Still, the AI is not the same as a real child. It has no body, emotions, or social needs, and it did not learn what the words actually meant. However, the study suggests that everyday speech contains enough patterns to support important parts of language learning. This AI may now give scientists a valuable new way to study how experience and memory work together during childhood.