PROSODIC MODELING IN SWEDISH SPEECH SYNTHESIS

Our present work concerns Swedish prosody in a speech synthesis framework. Two main problem areas are examined: prominence and phrasing. In a model for Swedish prosody, prominence levels (stress, accent, focus) are represented as layered and multidimensional for different domains (syllable, foot, word). Phrasing involves both coherence in the form of specific combinations of existing accentual gestures and separate boundary gestures. The main features of the intonation model are given in outline. Experiments on prominence include modelling of durations in a combined speech data base and rule synthesis framework, where the stressed-unstressed alternation appears to be the most important duration factor. Other experimentation concerns typical differences in the timing characteristics of the tonal gesture for focal accent between compound words and simplex accent II words. Experiments on phrasing include both production data from a varied speech material as well as synthesis and perception. Our experiments demonstrate that both coherence and boundary cues are effective as phrasing signals and that a combination of F0 and duration is typically used to signal phrasing. Our future plans include working with prosodic modelling of Swedish in a dialogue context and in a concept-to-speech framework.