PURPOSE: To perform a voice synthesizing process without losing delicate information that a natural sound source has and to obtain a voice of high quality by reading a sound source waveform table for continuous N-cycle waveforms (N: optional integer) in two directions.

CONSTITUTION: A sound source signal storage part l is stored with continuous N-cycle waveforms (N: optional integer) of sound source signals obtained by removing resonance characteristics of the glottis from the vowel stationary part of a natural voice previously. When a voice is synthesized, a two-way read sound signal generation part 2 reads the 1st cycle, 2nd cycle, etc., of the sound source signal out of the storage part l in increasing order from the head part of the waveform, reads the N-l-th cycle, N-2-th cycle, etc., in decreasing order after reading the waveform in the final Nth cycle, and then reads the cycles in increasing order again after reaching the 1st cycle. The two-way read is carried out continuously to generate a sound source signal with optional length. Then a fundamental frequency control part 3 converts the sound source wave outputted by the generation part 2 to a desired formant frequency and inputs it to a glottis filter part 4. Resonators corresponding to respective formant frequencies of desired phonemes vary the spectrum of the sound source signal to output the synthesized voice.