Interpolating Expressions in Unit Selection

In expressive speech synthesis, a key challenge is the generation of flexibly varying expressive tone while maintaining the high quality achieved with unit selection speech synthesis methods. Existing approaches have either concentrated on achieving high synthesis quality with no flexibility, or they have aimed at parametric models, requiring the use of parametric synthesis technologies such as diphone, formant or HMM-based synthesis. This extended abstract reports on on-going work exploring the addition of a certain degree of control over expressivity in a unit selection context. Rather than merely choosing one unit selection voice database in order to determine the expression contained in the generated speech, we use technology from the voice conversion domain to flexibly interpolate between two voice databases. This provides us with the possibility to generate a continuum of expressive tones between the two extremes defined by the two voice databases.