Science & Tech · AI · 14 hrs ago
AI models learned words on a schedule close to the children they heard
Princeton researchers trained AI models on recordings from eight babies’ homes in New Jersey.
The recordings followed the children for most of their waking hours across their first 1,000 days.
The models began with no dictionary, speech rules or text, and processed the audio in the order it was recorded.
They learned to identify recurring sound patterns linked to words as they heard people speaking around each child.
On average, each model reached halfway through a 649-word vocabulary checklist within 22 days of its paired child.
The children reached that halfway point at ages ranging from about 12 to 22 months, showing wide differences in when they learned words.
The study tests whether AI can learn from everyday experience over time, rather than from speech sorted into training batches.
Princeton researchers trained language models on recordings from eight children’s homes.
The models learned recurring sound patterns that corresponded to words.
Each model reached the halfway point on a 649-word vocabulary checklist within 22 days of the child it was paired with.
The recordings covered a median of 960 days per child, with between 1,900 and 6,600 hours of usable speech.
The models began with no dictionary, phonetics, word boundaries or text.
- Who
- Researchers at the Princeton Neuroscience Institute, led by Uri Hasson, studied eight children and trained a separate model on each child’s recordings.
- What
- Models learned recurring sound patterns corresponding to words on a schedule close to that of the children.
- When
- The recordings were collected from February 2022 to May 2025.
- Where
- The recordings came from eight homes in New Jersey.
- Why
- The researchers aimed to test whether a model could learn from a child’s actual experience over time.
This story does not have two clearly opposing sides.
Recording devices in the homes captured audio and video for 12 to 14 hours a day.
Caregivers filled out monthly checklists of words each child understood and said.
Each child’s usable speech amounted to 1,900 to 6,600 hours.
The models formed categories for nearly all English phonemes.
- Children studied
- Eight
- Vocabulary checklist
- 649 words
- Usable speech per child
- 1,900 to 6,600 hours
- Average model phoneme count
- 37.5
- Total project recordings
- Roughly 1.2 million hours of audio and video
What people are saying
We wanted to turn the usual logic of developmental research around
Instead of observing a child for an hour in the laboratory, we asked whether we could build a model capable of learning and immerse it in the child’s actual experience, day after day.