Feature Integration and Dimension Reduction in Unit Selection TTS

Unit selection based speech synthesis is able to generate high quality speech. During unit selection process, a set of unit candidates are provided for each target unit. Its performance is degraded with a large number of candidates. In order to improve the speed of the unit selection process, it is important to determine a small set of candidates. This needs a candidate selection method that does not miss good candidate units. In this work, a dimension-reduction based method is proposed to fuse different features together and use these features as the pre-selection criteria in the compilation of the candidate set in the unit-selection based speech synthesis method. Experiments in Mandarin TTS shows that speech quality can be improved using the proposed method.