Statistical Cues Facilitate Infants’ Discrimination of Difficult Phonetic Contrasts

Perhaps the best-known fact about developmental speech perception is that infants are remarkably adept at discriminating phonetic contrasts. In early infancy, this ability is unaffected by language environment, resulting in the surprising fact that infants can discriminate certain foreign contrasts that their parents cannot. For example, infants from English-speaking homes can hear the difference between [ž] and [ř], two fricatives used in Czech that Englishspeaking adults have difficulty discriminating (Trehub, 1976). Infants’ advantage at foreign contrast discrimination wanes over the course of the first year, though, as they gain experience with their native language; and by the age of 12 months infants no longer discriminate those foreign contrasts (Werker & Tees, 1984). Developmental speech perception, then, can largely be described as a process of paring down previously discriminable contrasts, to just that set of contrasts that is utilized in the native language. However, though infants’ discrimination of many phonetic contrasts exceeds adults’, there are in fact some phonetic contrasts that are difficult for infants. Among them are the contrasts between prevoiced and short-lag stops (Eimas, 1975; Aslin & Pisoni, 1980), between [d] and [ð] (Polka, Colantonio, & Sundara, 2001), and between fricative pairs such as [s]~[z] and [f]~[θ] (Eilers & Minifie, 1975; Eilers, 1977). Exposure to a language in which these contrasts are used phonemically apparently results in enhanced discrimination. For example, the [d]~[ð] contrast is discriminated well by adult English speakers, though not by speakers of French, a language in which it is not phonemic (Polka et al., 2001). The process of an infant’s developing perception of speech must therefore involve not only paring down of initially discriminable contrasts, but also enhancement of initially difficult contrasts. The goal of the present study was to examine how the facilitation of difficult native language contrasts might occur. In particular, the hypothesis we proposed was that changes in infants’ perception of difficult contrasts might be driven by statistical cues, based on the distribution of phonetic tokens in the input. Previous research has shown that phonetic contrasts are instantiated in the distribution of speech sounds produced in a language. For example, in Thai there are three voicing categories, while in English there are only two. This fact is evident from the distribution of stop consonant VOT values produced in each language: for Thai there is a trimodal distribution (the most commonly produced sounds form three clusters,