Preserving subsegmental variation in modeling word segmentation (or, the raising of baby Mondegreen)

Many computational models have been developed to show how infants break apart utterances into words prior to building a vocabulary—the “word segmentation task.” Most models assume that infants, upon hearing an utterance, represent this input as a string of segments. One type of model uses statistical cues calculated from the distribution of segments within the child-directed speech to locate those points most likely to contain word boundaries. However, these models have been tested in relatively few languages, with little attention paid to how different phonological structures may affect the relative effectiveness of particular statistical heuristics. This dissertation addresses this issue by comparing the performance of two classes of distribution-based statistical cues on a corpus of Modern Greek, a language with a phonotactic structure significantly different from that of English, and shows how these differences change the relative effectiveness of these cues. Another fundamental issue critically examined in this dissertation is the practice of representing input as a string of segments. Such a representation implicitly assumes complete certainty as to the phonemic identity of each segment. This runs counter both to standard practice in automatic speech recognition (where “hard decisions” are eschewed) and, more crucially, overestimates the ability of infants to parse and identify those segments from the spoken input. If even adult native speakers (with the benefit of higher-level linguistic knowledge, such as a