At the heart of generative phonology lies the assumption that the sounds of every language have abstract underlying representations, which undergo various changes in order to generate the 'surface' representations, that is, the sounds we actually pronounce. The existence, status and form of underlying representations have been hotly debated in phonological research since the introduction of the phoneme in the nineteenth century. This book provides a comprehensive overview of theories of the mental representation of the sounds of language. How does the mind store and process phonological representations? Kramer surveys the development of the concept of underlying representation over the last 100 years or so within the field of generative phonology. He considers phonological patterns, psycho-linguistic experiments, statistical generalisations over data corpora and phenomena such as hypercorrection. The book offers a new understanding of contrastive features and proposes a modification of the optimality-theoretic approach to the generation of underlying representations.