Evaluation of singing voice by amateur singers from various aspects
Akinori Ito · The Journal of the Acoustical Society of America · 2016
Many research works have been conducted for evaluating singing voice so far. Most of the works aimed to investigate the voice of professional vocalists. Besides, evaluation of amateur singers is useful for developing entertainment applications such as scoring of Karaoke. In this presentation, two of such applications are introduced: an evaluation of singing enthusiasm and the detection of mistakes of sung lyrics. The evaluation of singing enthusiasm is based on three kinds of features: the A-weighted power, the vibrato-related feature, and the fall-down feature. Using these features, perceived singing enthusiasm can be estimated relatively accurately. In addition, singers’ intention of enthusiasm can also be estimated from the fluctuation of the segmental power of the voice. Detection of mistakes of sung lyrics is based on the dynamic time warping between the input singing voice and the reference singing voice. To normalize the singer-dependent variation of voice features, we employed singer normalization based on linear regression. Based on the DTW alignment, frame-by-frame distances between the two voices are obtained. After temporal smoothing of the distances, the lyrics errors are detected based on the peak detection of the distance. High detection accuracy is obtained when the errors are based on phrases.