Generating script using statistical information of the context variation unit vector

Abstract

A statistical selection method is proposed for generating an optimized recording script for Concatenative Speech Synthesizer. This method starts with traveling a large text corpus to collect the statistical information of the Context Variation Unit Vectors (CVUV), which represent the multi-dimension phonetic contexts and properties of the synthesis unit. Each CVUV descriptor is organized as a node in a sorted tree of the CVUV forest to record the dimension values and the index to its position in the corpus. Then it selects sentences according to the pre-defined criteria relating to the CVUV distribution in the corpus. This selection algorithm has been implemented to generate syllable-based Chinese script and yielded satisfactory results. The context dimension definition concept is described in this paper, and the coverage analysis and computing time estimation are reported also.