Prosody generation in text-to-speech conversion using dependency graphs

The present paper addresses the problem of prosody assignment in the context of a system for text-to-speech conversion of Swedish. The question of what type of textual analysis is needed for prosodic purposes is discussed and the use of Dependency Graphs to this end is proposed. It is argued that this type of analysis provides sufficient syntactic information so that prosodic phrasing and prominence assignment algorithms can be successfully applied. The ability of the system to generate acceptable prosodic descriptions is evaluated by comparison with examples of prosodically analysed naturally spoken sentences.