To provide a voice signal processing device that can convert a singing voice signal of a singer to that of an original singer.

A pitch/sound volume extracting part 31 extracts pitch (frequency) data and sound volume data from a singing voice signal of a singer. An impulse generation part 32 generates a pulse train with an interval based on the pitch data. The pulse train has a wide frequency spectrum near a white noise, but a signal of a tone of an original singer is made by cutting this with formant of an original singer in a filter 33. This formant data is stored as sequence data with music data for KARAOKE playing, read out in parallel with KARAOKE playing, and written in a formant data buffer 36. A signal of a tone of an original singer can be made quite the same as that of a song of a singer by gain-controlling the signal of timre of an original singer with the sound volume data.