A corpus-based concatenative Mandarin singing voice synthesis system

A Mandarin singing voice synthesis (SVS) system is proposed in this paper. It generates a Mandarin song of an artificial singer based on the lyric and the music score information embedded in a MIDI file of the song. To get good quality of the song, two modules are presented, i.e., the synthesis unit selection module and the prosody and amplitude modification module. In the synthesis unit selection module, the corpus that complies with the lyric and closest to the music score information is selected. Then, an adaptive filter based prosody and amplitude modification algorithms are employed on the selected synthesis units. Through the proposed method, the system can synthesis any Mandarin singing voice on-the-fly by providing it the corpus of all syllables for male and female respectively. To increase the efficiency of the system, a preprocessing is also taken on the corpus. Finally, a subjective evaluation based on MOS is taken on the system and the synthesized sounds show good quality.