A simply configured speech synthesis device and the like for producing a natural synthetic speech at high speed. When data representing a message template is supplied, a voice unit editor (5) searches a voice unit database (7) for voice unit data on a voice unit whose sound matches a voice unit in the message template. Further, the voice unit editor (5) predicts the cadence of the message template and selects, one at a time, a best match of each voice unit in the message template from the voice unit data that has been retrieved, according to the cadence prediction result. For a voice unit for which no match can be selected, an acoustic processor (41) is instructed to supply waveform data representing the waveform of each unit voice. The voice unit data that is selected and the waveform data that is supplied by the acoustic processor (41) are combined to generate data representing a synthetic speech.