Daily replay helped AI learners discover thousands of words in Princeton study

An artificial learner exposed to a child’s everyday speech learned no words by the equivalent of the child’s second birthday. Giving it a daily chance to revisit recent recordings changed that: it discovered thousands. The contrast anchors a Princeton-led study exploring which learning mechanisms can help explain how children acquire language.

Published November 5 in Science, the study used everyday speech recordings from eight children, with a separate model learning from each child’s experience. The aim was to test an account of human development, not build a more capable chatbot. Princeton’s account describes a system that learned speech patterns without understanding what the words meant.

The recordings came from the 1,000 Days Project, which placed cameras and microphones in 15 homes from birth through roughly age three. Capturing most of each waking day gave the researchers a dense record of the speech surrounding individual children, rather than a brief observation in a laboratory.

Each model encountered speech in chronological order, one day after another. It received no text, dictionary or information marking where words began and ended. As the child’s waking hours grew, so did the model’s daily speech input. The researchers did not shuffle years of recordings into a single training set.

The team began with recordings from a child given the pseudonym Coral. Simply moving through that stream of speech produced little lasting learning: the model retained almost none of its experience from one day to the next.

The researchers then borrowed an idea from sleep: replaying recent experience to help retain it. At the end of each simulated day, the model revisited a limited sample of speech, mostly from the preceding few days. This was not an attempt to reproduce sleep’s stages or cycles, only one proposed memory-building mechanism.

With replay, Coral’s model first identified units matching nearly all English speech sounds, known as phonemes, then began detecting recurring words. By the simulated second birthday, its vocabulary numbered in the thousands. Using the same settings, models trained on speech from the other seven homes also discovered English sounds and thousands of words.

The households differed in how much speech children heard, who spoke to them and what families discussed. Yet the models progressed at roughly the same pace. That consistency was also a mismatch: parents reported substantially different vocabulary timelines for the children, with some reaching a large vocabulary nearly a year before others.

To capture those differences, the team changed the learners rather than their recorded environments. Three settings varied between models:

• How much recent speech each model replayed each night.
• How quickly it learned from the speech it encountered.
• When it began learning words.

Adjusting those settings allowed each model to reproduce the timing of the child it represented, the researchers reported. The same starting learner could succeed across homes, but matching individual developmental trajectories required differences in the learning process itself.

Unlike much of modern AI, which aims at superhuman performance, we are trying to understand human learning and how it gives rise to human intelligence

Hadas Raviv, computational cognitive scientist who led the study’s computational modeling

The researchers emphasize that the models are not children. They have no bodies, social motivations or developing brains. They cannot point to an object, watch a parent’s face or use language to attract attention. Most importantly, they do not know what any word means.

The team presents the framework as a way to test developmental theories by building learners, exposing them to a child’s actual experience and comparing their progress. Hasson said the longer-term goal is to give models more capacities reflected in the recordings, including seeing, interacting and acting in the world.