How do infants access language? Listening to us talk!

Language acquisition is complex. Depending on the countries of the world and the languages ​​spoken there, all children do not learn the same thing. Take the example of Japanese for example, which does not distinguish between the sounds “R” and “L’ as in “roc” or “foul”. Thus, one-year-old babies, for example, less easily distinguish the sounds “roc” and “foulbrood” when they live in an environment where Japanese is spoken rather than French. This is what researchers from the University of Maryland (United States) have noticed. Study results were published in Proceedings of the National Academy of Sciences.

Groups of words that don’t work

For the scientific community, this phenomenon of early phonetic learning of babies is based on the fact that they group sounds into phonetic categories of the vowel and consonant type by a mechanism known as distributional learning. However, the researchers of this study have doubts about this mechanism, which prompted them to try the experiment.

“Assumptions about what is learned by infants have traditionally guided researchers in their attempts to understand this surprising phenomenon.says Thomas Schatz, a postdoctoral associate at the University of Maryland’s Institute for Advanced Computer Studies and lead author of the study. NOTWe propose to start from hypotheses about how infants might learn.”

To find out for sure, the researchers introduced a quantitative modeling framework based on a large-scale simulation of the language learning process in infants. This approach, which uses machine learning techniques, makes it possible to systematically link the mechanisms of learning to verifiable predictions concerning the adaptation of infants to their mother tongue.

An unsuccessful method

To do this, they therefore designed an algorithm capable of simulating the learning process of infants. Trained in English and Japanese through voice samples, it was possible to better understand how a child acquires his first phonetic knowledge. Once done, two questions remained unanswered: Could they explain the observed differences in how infants learn Japanese and English by discriminating speech sounds? Above all, do the models really discriminate between the phonetic categories of vowels and consonants?

For the first question, their models effectively took into account the observed behavior of infants, in particular the difficulty of Japanese infants to distinguish certain words because of the sounds that are too close (the “R” and the “L” in this case) . On the other hand, contrary to what was thought, it is impossible for machines (which are modeled on the learning rhythm of children) to distinguish vowels from consonants. This therefore calls into question all the pre-existing knowledge on early phonetic learning.