Early phonetic learning without phonetic categories--Insights from machine learning

Decades of work have shown that even before they can talk, human infants are quickly tuning into the properties of the language (s) spoken in their environment. By their first birthday, they have already become attuned to the sounds of their language (s), processing native phonetic contrasts more easily than non-native ones. For example, between 6-8 months and 10-12 months, infants learning American English get better at distinguishing American English [ɹ] and [l], as in ‘rock’vs ‘lock’, relative to infants learning Japanese. This phenomenon has been dubbed early phonetic learning and has been taken as evidence that, during the first year of life, infants form phonetic categories like [ɹ] and [l]. We present evidence that calls this interpretation into question. We show that a machine learning model trained without supervision on raw, unsegmented speech recordings can predict the documented changes in discrimination of [ɹ] and [l] in Japanese and American English infants—but that it does so by learning units that are too localised in time and acoustically variable to correspond to phonetic categories. These results constitute the first demonstration of a feasible mechanism for early phonetic learning under realistic learning conditions. They help resolve a tension between the hypothesis that early phonetic learning is driven by statistical analysis of the speech sounds in the infant’s environment, and the difficulty encountered by computational models to robustly discover phonetic categories using such mechanisms. More broadly, our results challenge the view commonly held for several decades that perceptual changes in infancy constitute …