Skip to the cards
Subject
Nutrition
Filed
Source
Shelly Fan
Loans
735

AI Models Learned Early Speech Patterns From Children’s Daily Recordings

Models trained on recordings from eight children learned speech sounds and early words, though replay and model tuning were needed to match development.

AI Models Learned Early Speech Patterns From Children’s Daily Recordings

Models trained on recordings from eight children learned speech sounds and began recognizing words before age two. The result suggests that language learning can be studied through a child’s everyday experience, but it doesn’t show that an AI learns as a baby does. The models needed repeated exposure to recent sounds, and they heard speech without the gestures, faces and social cues that shape children’s learning.

The Princeton University team drew on recordings gathered through the First 1,000 Days project. Fifteen families took part, with different backgrounds, routines and language environments. Cameras and microphones captured more than a million hours of audio and video across the homes. For each child, the researchers selected up to 6,600 hours of audio for model training; the reported learning results cover eight children.

Learning from one child’s experience

Each model began without online information, built-in English sound categories or labels marking where words began and ended. Its audio encoder divided speech into 20-millisecond segments. A predictive decoder then learned to anticipate the sounds that might come next.

The sequence of exposure mattered. Instead of giving a model all its recordings at once, the researchers fed it a child’s sound environment over time, day by day. The aim was to approximate the changing stream of speech a child encounters while growing up.

That approach alone produced weak results. The team found that even thousands of hours of speech from one child weren’t enough for the model to learn well through simple exposure. They improved learning by adding a daily replay period, during which each model revisited snippets of recent speech. The researchers drew inspiration from the role of sleep in consolidating memories, though replay in a model is not the same as sleep in a child.

With replay, the models moved from identifying phonemes—the basic sound units of speech—to recognizing words that appeared repeatedly. Across the eight children, models learned about 37 phonemes within 15 months and began recognizing words before age two. Learning picked up pace over time, much as children’s word recognition did in the comparison described by the researchers. By roughly 1,000 days, the models had learned thousands of words.

The pace varied widely from child to child. The team adjusted settings, including how much speech each model replayed, to better match an individual child’s learning path. That tuning helped align model performance with observed development, but it also means the results don’t establish one fixed recipe that explains how every child learns language.

What infant research already shows

The work builds on decades of evidence that babies detect patterns in speech well before they can explain or repeat what they hear. In a study from the 1970s, one- and four-month-old infants changed their sucking patterns when they heard a new speech sound, such as the difference between “b” and “p.” By eight months, babies can pick out words from a continuous stream of artificial speech even when there are no pauses to mark word boundaries.

Research tracking infants’ eyes has also found rapid growth in spoken-word recognition between 15 and 25 months. More recent studies have used microphones and wearable recorders to connect children’s emerging vocabulary with the audio and social interactions in their daily lives. Computational models have helped map parts of language development, though many have focused on specific skills. The Princeton study adds a longer, more individualized record of children’s sound environments.

Collecting that kind of record raises questions about privacy. The team created a framework that let parents retain ongoing control, including the option to withdraw or delete recordings. The households used between four and 14 recording devices, which ran for about 12 hours a day. An automated system identified when a baby appeared on camera and detected speech; researchers removed noise, silence and overlapping conversation before preparing audio for model training.

A useful model, with clear limits

The study offers a way to test ideas about language learning against recordings of children’s actual environments. It may also help researchers examine what happens when learning develops differently. But the models are not digital copies of children’s brains: they don’t process information through the parallel, energy-efficient activity of neurons, and the models in this study learned from sound alone.

Babies also use gaze, facial expressions, gestures and exchanges with other people to make sense of what they hear. The recordings include video, but those visual and social signals weren’t part of the models’ language learning here. Researchers’ next task is to work out which parts of the models’ success reflect children’s learning processes and which come from computational shortcuts.

Adding information about what children see and how they interact could make future models more realistic, while helping researchers ask sharper questions about early development. For now, the main finding is narrower: a model exposed to one child’s speech over time could learn early sound and word patterns, but only after daily replay helped it retain recent experience. The work provides a tool for studying language development; it doesn’t settle how children acquire language or show that current AI learns the way they do.

Source: Shelly Fan · singularityhub.com

Filed alongside