Modelling speech prosodics for synthesis-perspectives and trials

Our paper is about syllable-based synthesis and prosodics. It considers perspective carefully and suggests that a particular way of looking at the overall communication system (speaker and perceiver together) prompts a special approach to speech production. We use this idea to give examples of a synthesis strategy in the prosodics area resulting in demonstrably improved naturalness. Our perspective in approaching speech synthesis is to set as goal the creation of a good percept in the mind of the listener-a percept which truly represents the speaker´s plan and which minimises the work the listener must do to achieve the right repairs to our relatively poor synthetic signal. By systematically observing listeners´ reactions to progressively improved or degraded signals we are gradually building a model of the repair processes. The idea is to introduce into our synthetic speech special repair-oriented cues which, we hope, will cause the listener to report an improvement in the quality of the signal, although the signal itself is no closer to a human signal than it was before. What we are doing is creating synthetic speech to be perceived, not synthetic speech which is to be tested against a natural signal. While we wait for a perfect production model geared to the needs of synthesis, we feel that we can make significant progress with this alternative approach