Singing Voice Analysis for Singer Identification Using Vibrato Features Deepali Y Loni* and Dr Shaila Subbaraman** *Department of Electronics, Textile & Engineering Institute Ichalkaranji, India deepaliloni@rediffmail.com **Ex-Academic Dean, Walchand College of Engineering Sangli, India shailasubbaraman@yahoo.co.in Abstract: Vibrato is one of the key acoustic feature of singing voice. This work investigates the measurement of vibrato by extracting it from natural cappella section of singing voice considering variety of singing styles and tones. The algorithm performs detection, extraction and analysis of vibrato parameters; rate and extent. These vibrato parameters are explored for singer identification and is validated on a database of 11 singers (5 female and 6 male singers) containing more than 50 cappella segments from several songs of each singer derived from commercially available CD recordings. A maximum accuracy of 62% is achieved in identifying the singers. This indicates the significance of vibrato as an important characteristic of singing voice. Keywords: Vibrato, cappella, pitch contour, singer, confusion matrix. Introduction One of the widely used quality measuring benchmark of singing is vibrato. Vibrato is the skill acquired by the singers after many years of extensive vocal training. The vibrato adds certain naturalness to the singing voice and it is a very specific characteristic of the singing voice of a singer [1]. It is thus considered as one of the distinct factor that differentiates singing from plain speech. Vibrato is defined as periodic variations of the fundamental frequency of the singing voice around an average value. “Vocal Vibrato” is also said to be the ornament of singing voice. Singers use vibrato to enhance their expressiveness and to place emphasis on significant words or phrases of a musical piece [2]. Major of the research work in this direction has revealed that trained singers have dominant presence of vibrato in the singing segments as compared to untrained singers [3]. Investigations have also revealed that vocal training improves vibrato as well as other acoustic features like formant and perceived quality of the singing voice [4]. The natural production of vibrato sound is the result of rhythmic variation of larynx muscles in response to sub-glottic pressure [5] which indirectly is singer specific characteristic and can be explored for singer identity. Edward [6] found that not all singers present the same quality of vibrato in terms of deviation of vibrato confidence. The results presented in [7] revealed the variations in the vibrato of the professional singers and also indicated that when a singer imitates another singer, the vibrato features are consciously controlled. The effectiveness of vibrato-motivated feature; octave frequency cepstral coefficients (using triangular, parabolic and cascaded subband filters) is explored by Tin Lay et. al. [8] in order to identify singers of popular music. The proposed work explores vibrato characteristics of different singers. Major of the earlier studies extract vibrato by making the singers sing the same tone in rather slow tempo to produce large number of vibrato cycles. In contrast to these set conditions, the proposed investigation analyzes the vibrato information of natural singing voice by identifying the vibrato sections present in different segments of a song. The vibrato parameters; rate and extent are computed to analyze and identify the singer. Extracting Vibrato Sections from Singing Voice The most important characteristics that define the vibrato acoustic are the vibrato frequency (measured in Hertz) and the vibrato extension (measured in semitones). The vibrato frequency is a measurement of the speed of pitch variation; commonly called as vibrato rate. The vibrato extension also known as vibrato extent is the measurement of vibrato depth. A typical vibrato segment extracted from a singing voice is as shown in Fig. 1.