Get the app

This AI Learned Language From the Recordings of a Baby’s First 1,000 Days

Singularity Hub
Read full story· 6 min
Princeton researchers created AI models that learn language by mimicking babies' real-life auditory experiences over their first 1,000 days, differing from traditional large language models that train on vast text datasets. This approach aims to uncover how infants acquire language naturally and inspire more efficient AI learning methods.

What happened

Researchers at Princeton University developed AI models that learn language by processing real-life audio recordings from the first 1,000 days of eight children's lives. The recordings were collected from cameras and microphones installed in the children's homes, capturing their daily auditory environments. Each AI model was trained to experience the language input as the child did, day by day, rather than all at once. The project aimed to better understand how infants naturally acquire language, using a method different from traditional AI models that rely on large text datasets.

Key details

  • Recordings were collected from 15 families, with 4 to 14 devices per home, for about 12 hours daily.
  • Over a million hours of audio and video were captured, with up to 6,600 hours of audio used per child.
  • The AI models started with no prior knowledge of language or English sounds.
  • Replay of daily speech was added to improve the AI's learning, inspired by how human memory consolidates during sleep.
  • Across eight children, the models learned about 37 phonemes in 15 months and began recognizing words before age two.

Why it matters

This research affects AI developers and scientists studying language acquisition by providing a new method for training models with less data and energy than traditional approaches. Families and children participating in the study contributed to a dataset that could help design more efficient and human-like language learning systems. The findings may lead to AI that requires less computational power and could offer insights into how language learning can go wrong in early childhood.

#child-development#language-acquisition#princeton-university

More in AI Research