Towards Automatic Measure of Similarity for Use in Unit Selection

The present paper focuses on the unit selection approach to speech synthesis, discussing drawbacks mainly related to the current handling of target features that basically results in the need of huge corpora. In the paper there are outlined possible solutions based on measuring (dis)similarity among prosodic patterns. In the initial experiment, trying to verify the feasibility of the proposed solution, the (dis)similarity of acoustic signal measured by different techniques is correlated to perceived similarity estimate obtained from a large-scale listening test.