PURPOSE: To obtain a synthesized speech which is smooth and close to a natural voice.

CONSTITUTION: A waveform storage device 17 storeps groups of speech waveforms which are the closest to the centers of gravity of respective groups and selected in a speech data base 13 by clustering frequency spectra of a speech signal as feature parameters, spectra (reference spectrum) in the centers of gravity, and pitch marks of respective speech waveforms as a gathering of speech units, and, an input text is made into a phoneme series, speech waveforms, reference spectra, and pitch marks corresponding to the respective phonemes are selected from the waveform information storage device 17, and a spectrum characteristic varying circuit 36 varies the spectra so that they approximate a standard spectrum. Then, the varied waveforms are supplied to a waveform synthesizing circuit 23 to control pitch and power continuing length as usual, thereby synthesizing a speech.