Refining Unit Boundaries for Mandarin Text-to-Speech Database

Abstract

In unit selection based Text-to-Speech (TTS) synthesis, the accurate position of the unit boundaries in the unit selection database is one of the factors that determine the quality of the synthesized speech. To ensure the accuracy of the boundary positions, developers often have to manually verify the speech boundaries that are generated by automatic speech recognition techniques. In order to reduce the manual workload, it is necessary to use automatic methods of refining the position of the unit boundaries. This paper proposes a frame-shift method to find the globally optimal joint position for unit concatenation between any two matching units. Experiment results show that this method can improve the boundary accuracy compared to manual labeling.