Speech synthesis in dialogue systems

This paper argues that dialogue synthesis has special requirements. Users expect a high level of naturalness. We argue for good segmental rendering with the addition of pragmatic effects, such as emotion and attitude. We discuss the language model used, emphasising the phonological and phonetic levels essential to handle these effects. We draw on the theories of Cognitive Phonetics and Pragmatic Phonetics. INTRODUCTION Dialogue involves a human user and a computer interacting to exchange information. A telephone inquiry system is an example of dialogue based interaction for requesting and providing information. Dialogue systems incorporate several identifiable elements including a dialogue controller [1] and input/output subsystems taking the form of automatic speech recognition and speech synthesis. In this paper we address a number of questions determining how to provide a synthesis module to meet the demands of dialogue. In particular we consider € general prerequisites of synthesis, € special requirements for synthesis in dialogue systems, € modelling the requirements in the dialogue context, € implementation of the model. We make the assumption here that other components of the overall system have determined such matters as € what is to be spoken, € how it is to be spoken. However, the implementation of the output will determine the formulation of these other components of the system. GENERAL PREREQUISITES To ensure user acceptability in a wide range of environments the speech produced by a synthesiser must be of high quality. Almost all synthesisers now produce highly intelligible output, and listeners have little difficulty in determining the intended message. Although there are no universally accepted formal evaluation procedures for synthetic speech, there is little disagreement that generally the threshold of good intelligibility has been passed. Much of the achievement is due to improvements in the way the sound wave itself is generated. To address the question of general requirements satisfactorily we need to make a distinction between low level synthesis and high level synthesis.