A Study of Vibrato Features to Control Singing Voices

Singing synthesis applications have recently been developed to create natural singing voices. Applications such as VOCALOID can synthesize singing voices by inputting lyrics and scores, and musical expressions such as vibrato and portamento are added to the singing voices by controlling pitches. This control is crucial to synthesize natural singing voices, but users are required to edit the pitch contour precisely, which means that the sound quality depends on the skill of the user. If the singing voices were controlled by using professional singers’ musical expressions, users could synthesize natural singing voices more easily. We created a singing database composed of several singers’ vibrato and portamento and then extracted their musical expressions and analyzed the vibrato to synthesize natural singing voices. The database consists of four professional singers (two females and two males) who sang the five Japanese vowels. The vibrato pitches were set to thirteen pitches of an octave that each singer could sing. This database contains not only the singers’ own singing voices but also imitations of other singers. We used the database to analyze the differences of vibrato between professional singers and to determine how to control vibrato by impersonation. In a previous study, a vibrato model in the form of a fundamental frequency (F0) contour was proposed to synthesize natural singing voices. This model can control the vibrato rate (speed of pitch fluctuation) and vibrato extent (width of pitch fluctuation). Vibrato sections are extracted and vibrato features are calculated from the F0 contour of singing voices and defined as average in the vibrato section. It was reported that these vibrato features varied according to the type of song. Song type was the only focus of this study, whereas in our study, we focus on the difference of vibrato based on the singer. We propose using vibrato duration (the ratio between vibrato length and voice length) and time fluctuations of the vibrato rate and vibrato extent to analyze professional singers’ vibrato. The time fluctuation of the vibrato rate is calculated from period of F0 contour, and the time fluctuation of the vibrato extent is calculated from the instantaneous amplitude of the F0 contour. Results showed the different vibrato features of the professional singers and also indicated that when a singer imitates another singer, the vibrato features are consciously controlled.