Attributes of audio feature contours for automatic singing evaluation

This paper concerns the automatic singing assessment by use of the audio feature contours. We show that statistical properties of feature trajectories can be exploited to create voice-training system to judge the quality of singing. We have analysed a large set of audio features, using three feature selection methods, in order to determine descriptors with high discrimination power. Obtained feature vectors are fed up into support vector machine (SVM) classifier with RBF kernel for classification of singing phrases. In the result, singing phrase is assigned to one of three quality categories. The proposed approach yields to comparable assessment accuracy with a group of voice experts, where taking specific properties of training phrases into account can improve the final accuracy.