A sentence or a singing is to be synthesized with a natural speech close to the human voice. To this end, singing metrical data are formed in a tag processing unit 211 in a singing synthesis unit 212 in a speech synthesis apparatus 200 based on singing data and an analyzed text portion. A language analysis unit 213 performs language processing on text portions other than the singing data. As for a text portion registered in a natural metrical dictionary, as determined by this language processing, corresponding natural metrical data is selected and its parameters are adjusted in a metrical data adjustment unit 222 based on phonemic segment data of a phonemic segment storage unit 223 in the metrical data adjustment unit 222. As for a text portion not registered in the natural metrical dictionary, a phonemic symbol string is generated in a natural metrical dictionary storage unit 214, after which metrical data are generated in a metrical generating unit 221. A waveform generating unit 224 concatenates necessary phonemic segment data, based on the natural metrical data, metrical data and the singing metrical data to generate speech waveform data.