There are several areas of text-to-speech woik which have often been neglected by researchers, but which become vitally important once an application for a system is envisaged. This paper highlights some of these areas, concentrating on text normalisation and pronunciation. Examples are given of areas being addressed in the text-to-speech system currently being developed at BT Laboratories. Our data-driven approach to developing pronunciation methods, particularly for proper names is explained. An overview of the BT Laureate text-to-speech system is given showing the general structure and, in particular, where text normalisation and pronunciation fit into the structure.

Keywords: text normalisation, pronunciation, grapheme to phoneme alignment, pronunciation by analogy