TABLE OF CONTENTS Chapter written by Olivier BO FFARD and Christophe D'ALESSANDRO.
Nowadays, the telecommunication sector is experiencing a strong growth in services proposed to clients by large telephone operators. Ultimately, these services rely on the transmission of a voice signal over a network and require ergonomic interfaces that are becoming increasingly complex and more and more similar to human interaction. A man-machine dialog engine coupled to speech recognition and spoken message generation systems constitute the technological backbone of the vast majority of existing systems handling spoken information.
Most voice servers replay pre-recorded messages [1] spoken by a person who has lent his or her voice to the application beforehand. From the service provider's viewpoint, two categories of services can be distinguished: firstly, services for which the entire message has been stored as an acoustic signal, possibly in a compressed form, and which is generally replayed at 32 or 64 kbit/sec; secondly, services delivering messages which result from the concatenation of fixed portions and variable parts. The latter approach is slightly more flexible than the former and offers a wider range of messages for the application. While the quality of the messages are very close to those of natural, spoken utterance, it is difficult to imagine such a solution being tractable for large vocabulary applications (for example, reverse directories), because storage costs and the expenses involved in updating the inventory of recorded names would clearly be prohibitive.
A text-to-speech (or TTS) synthesis system aims at generating a speech signal that corresponds...
