A voice analysis method comprises generating a time series of a relative pitch (R), which is a difference between a pitch (PB) generated from music track data (XB) designating respective notes of a music track in time series, and a pitch (PA) of a reference voice. The music track is divided into unit sections (UA) of a predetermined duration, and singing characteristics data (Z) is generated, which includes, for each of a plurality of statuses (St) of a model (M), classification information for classifying the unit sections (UA) into a plurality of sets and variable information defining a probability distribution of the time series of the relative pitch (R) within each of the classified unit sections (UA). The classification information is generated based on a condition relating to an attribute of the note and based on the condition relating to an attribute of the each of the unit sections (UA).