Early phonological acquisition: computational and experimental approaches

old_uid2471
titleEarly phonological acquisition: computational and experimental approaches
start_date2007/03/20
schedule15h45
onlineno
location_infoBig conference Room (1.63)
summaryDuring the first year of life, infants rapidly acquire many aspects of the sound structure of their native language, such as its vowel and consonant inventories, and constraints on the shape of its words. In this talk I concentrate on the acquisition of a more abstract property, i.e. phoneme categories. Phonemes are abstract units that can have several phonetic realizations. An acquisition algorithm that does not take lexical knowledge into account is based on the observation that sounds that are realizations of a single phoneme do not occur in the same contexts, hence, have complementary distributions. I will show that a statistical version of this algorithm allows sounds to be grouped into phoneme categories even in the presence of noise. Using simulations on phonetically transcribed speech of adults addressed to infants, I also show that this algorithm is too powerful and generates numerous false alarms, consisting of pairs of sounds that have (near-)complementary distributions but that are realizations of different phonemes. A filtering mechanism that is based on linguistic properties of phoneme categories and that uses either an articulatory or an acoustic distance metric will be shown to effectively discard these false alarms. Finally, I will present results from a series of perception experiments showing that 12-month-old infants are sensitive to complementary distributions of sounds. Together, these results suggest that infants could use the algorithm and hence group sounds into abstract phoneme categories before having much lexical knowledge, provided they know linguistic constraints upon the nature of these categories and can use an articulatory or acoustic distance metric.
responsiblesZondervan