Detail publikace

Abstrakt

This paper describes the issues of the usage of various phase component types in the development of the Czech TTS system based on harmonic sinusoidal signal representation. We have found the approaches for speech representation based on sinusoidal coding [1] or harmonic plus noise modelling [2] very promising. It is mainly due to possibility of high compression of the spectral representation and possibility to 'smooth' the transitions on the spectral level. The major inconvenience is the necessity to use natural phase components to reach quality synthesis with preserved naturalness. Trying to interpolate the phase components across the concatenations causes the discontinuities in generated signal. We found that the discontinuities substantially degrade the fluency of synthesized speech. We propose the method of substituting the phase components by one locally constant phase component to guarantee the local phase coherence.