Modelling the Evolution of Speech Segmentation
- Richard Littauer, University of Edinburgh
- Old College Lorimer Room, University of Edinburgh
Recent studies have shown that 8-month-olds can segment continuous strings of speech
syllables into word-like units using only statistical computation of syllables, without relying on
acoustic or prosodic cues for word boundaries. (Aslin et al. 1997, 1998; Mattys et. al, 1999) These
studies looked at phonotactic regularities and syllable transition probability, but did not take into
account different types of statistical processes. In this study, I used a computational simulation to
examine four different statistical processes that may be used to parse uninterrupted phonological
strings: word recognition, word frequency, syllable transition count, syllable transition probability
(as well as a priori word retention and recognition). Each of these four techniques is judged by
analysing word pairs against a pre-existing syllable string created using a limited lexicon. The word
pairs contain one of the original words and a scrambled, random, or chopped word. The more
probable word is put into a list used to create the next string; By filtering the output of these
processes through an Iterated Learning Model over hundreds of generations, the aptitude of each
different statistical process to explain syllable segmentation in infants can be ascertained. This
study will present a more complete picture of not only syllable segmentation but also the evolution
of lexical discreteness and phonotactic rules, by analysing the trends in transitional probabilities
word-internally for each generation. I will present my findings, which will be published in my
forthcoming dissertation.
syllables into word-like units using only statistical computation of syllables, without relying on
acoustic or prosodic cues for word boundaries. (Aslin et al. 1997, 1998; Mattys et. al, 1999) These
studies looked at phonotactic regularities and syllable transition probability, but did not take into
account different types of statistical processes. In this study, I used a computational simulation to
examine four different statistical processes that may be used to parse uninterrupted phonological
strings: word recognition, word frequency, syllable transition count, syllable transition probability
(as well as a priori word retention and recognition). Each of these four techniques is judged by
analysing word pairs against a pre-existing syllable string created using a limited lexicon. The word
pairs contain one of the original words and a scrambled, random, or chopped word. The more
probable word is put into a list used to create the next string; By filtering the output of these
processes through an Iterated Learning Model over hundreds of generations, the aptitude of each
different statistical process to explain syllable segmentation in infants can be ascertained. This
study will present a more complete picture of not only syllable segmentation but also the evolution
of lexical discreteness and phonotactic rules, by analysing the trends in transitional probabilities
word-internally for each generation. I will present my findings, which will be published in my
forthcoming dissertation.