The "soul" of language does not use statistics: reflections on vowels and consonants

This paper reviews studies of language processing with the aim of establishing whether any type of statistical information embedded in linguistic signals can be exploited by the language learner. The constraints as to the information that can be so used, we will argue, should be used to inform theories of language acquisition. We present two experiments with their respective controls. Both show that consonants (Cs) are much more suitable than vowels (Vs) to parse speech streams using statistical dependencies. These experiments use streams composed of items in which statistical information is carried either by the sequence of consonants or by the sequence of vowels. Both kinds of items are simultaneously present is the speech stream but, crucially, their overlap is only partial. Since the location of dips in transitional probabilities (TPs) between adjacent syllables differ for the first and the second type of sequences, we can explore whether consonants and vowels are equally efficient segments to parse signals. Our results show that "consonant words" (CW) are significantly preferred over "vowel words" (VW). We discuss the implication of our results for models of language acquisition.