The development of phonological representations remains a hot topic within both the developmental and neural network literature. Historically, theoretical accounts have fallen within one of two camps: the accessibility account which proposes that phonological representations are adult-like from infancy (Rozin & Gleitman, 1977; Liberman, Shankweiler & Liberman, 1989) and the emergent account which proposes that phonological representations become gradually restructured over development (Metsala & Walley, 1998; Ventura, Kolinsky, Fernandes, Querido & Morais, 2007; Ziegler & Goswami, 2005). Within this thesis we tested predictions made by the accessibility account and key variants of the emergent account using data from both behavioural (Chapters 2, 3 and 4) and neural network studies (Chapter 5). The novel measures used within Chapters 2 to 4 were devised to allow us to contrast implicit measures of phonological representation (PR) which probe the segmentedness of the representations themselves, with explicit PR measures which tap into children's conscious awareness of phonological segments. Within Chapter 2 we present evidence that while explicit awareness of phonological structure is dependent on letter-sound knowledge, implicit sensitivity to the segments within words emerges independent of literacy. Within Chapter 3 a longitudinal study investigated the segmentedness of children's phonological representations at the rime and phoneme level. These results demonstrate that implicit sensitivity to both rime and phoneme segments is driven by vocabulary growth and is not dependent on letter-sound knowledge. The results within Chapter 3 also suggest that, while awareness of rime segments emerges naturally through oral language experience, explicit awareness of individual phonemes is related to letter-sound knowledge. In Chapter 4 we explored the idea of global versus phonemic representation using a mispronunciation reconstruction task. We found that sensitivity to both global and phonemic similarity increased over time, but with global sensitivity reaching adult levels early on in development. In Chapter 5 a neural network was trained on the mappings between real acoustic input and articulatory output data allowing us to simulate the development of phonological representations computationally. The simulation data provide further evidence of a developmental increase in sensitivity to both global and phonemic similarity within a preliterate model. Taken together, the results provide strong evidence that as children's vocabularies grow they become increasingly sensitive to both the global properties and segmental structure of words, independent of literacy experience. Children's explicit awareness of phonemes, on the other hand, seems to emerge as a consequence of learning the correspondence between letters and sounds. Within the context of the wider literature, the current results are most consistent with the PRIMIR framework which predicts early detailed phonetic representations alongside gradually emerging phonemic categories (Werker & Curtin, 2005). This thesis underlines the importance of using implicit measures when trying to probe the representations themselves rather than children's conscious awareness of them. The thesis also represents an important step towards modelling the emergence of segmental representation computationally using real speech data.