There are provided a singing voice-synthesizing method and apparatus capable of performing synthesis of natural singing voices close to human singing voices based on performance data being input in real time. Performance data is inputted for each phonetic unit constituting a lyric, to supply phonetic unit information, singing-starting time point information, singing length information, etc. Each performance data is inputted in timing earlier than the actual singing-starting time point, and a phonetic unit transition time length is generated. By using the phonetic unit transition time, the singing-starting time point information, and the singing length information, the singing-starting time points and singing duration times of the first and second phonemes are determined. In the singing voice synthesis, for each phoneme, a singing voice is generated at the determined singing-starting time point and continues to be generated for the determined singing duration time.