AUTOMATIC SEGMENTATION FOR CONSTRUCTION OF SIGNAL DICTIONARY IN CONCATENATIVE SYNTHESIS

摘要

The quality of a concatenative text-to-speech synthesis (TTS) system is directly related to the accuracy with which the underlying acoustic inventory is formed. Preparing those inventories is very tedious and time consuming. Here we proposed an automatic segmentation procedure for the generation of these signal units, based on the extraction of some signal domain parameters. An optimized histogram process is used to get primary segment positions, which are over-segmented in nature. Finally, using the generalized slopes of the parameters, the unnecessary segment markers are removed.