From speech synthesis by analogy to simulations of human reading

Text-to-speech synthesis is concerned with automatically determining the pronunciation of words from their spellings. This field has important applications including telecommunications-based information services and aids for blind and non-speaking people. There is an apparent commonality between the problem of speech synthesis and the process of human reading. Given this situation, it is surprising that there has been minimal interaction between the two fields. On one hand, studies on human reading can bring some valuable insights into the very challenging problem of mapping letters to sounds in speech synthesis. On the other hand, the study of human reading can benefit from the precision of mathematical models. In the last few years, the applicant and colleagues have developed a method for converting letters to sounds for text-to-speech systems. The research program is concerned with using this computer method as a way of increasing our understanding of how people read aloud. This proposal can be divided into three main objectives: 1) Improving our existing model for text-to-speech applications; 2) Demonstrating the capability of our model to account for the key human reading phenomena; 3) Extending the model to languages other than English, as a way of furthering the understanding of the relation between spelling and speech sounds.