| 摘要 |
PURPOSE:To improve the naturalness of a connection between CV data and VC data by performing data generation and storage wherein connection points of a consonant plus a vowel (CV data) and a vowel plus a consonant (VC data) are equalized in energy. CONSTITUTION:A speech waveform is fetched, syllable by syllable, its level is normalized, and a speech parameter is analyzed to perform parameter extraction. The parameter is made into time-series data and extracted as energy and pitch frequency, and when CV data and VC data are determined corresponding to the parameter, data segmentation wherein the energy levels of the end of the CV data and the start of the VC data in the time-series data are added is performed. The CV data and VC data after the data segmentation are stored on a syllable parameter memory as respective parameters. Therefore, when a speech is synthesized by connecting the CV data and VC data, abrupt variation in energy level at the connection point of the CV data and VC data is reduced to obtain natural high energy variation, thereby improving the naturalness of the synthesized voice. |
