Analysis of Audio Descriptor Contribution in Singer Identification Process

An audio descriptor describes the information of an audio signal in a compact and precise representation. There are various standards available to extract the audio information in various ways to be used for particular applications such as, speaker recognition, musical instrument identification, singer identification, multimedia database indexing, genre detection and so on. There are various audio descriptor extraction and manipulation tools available to be used with MatLab. Audio descriptor plays vital role for the applications such as singer identification. The process becomes complex with respect to monophonic, homophonic and polyphonic type of music. In North Indian classical music, (homophonic version), an accompanying instrument called Tanpura is continuously played during the vocal performance of the singer. In such cases, merely standard procedures of speaker recognition or singer identification from polyphonic music are not sufficient. The contribution of different audio descriptors changes from one type of input to the other. In this paper, we have established a singer identification system using MIRtoolbox for Timbral feature extraction and KMeans clustering for classification. The purpose of this paper is two folded. A) to design and implement a singer identification system and B) to find the contribution of the audio descriptors from each level to the final accuracy of singer identification result. With the method proposed, an accuracy of 96.6667% is achieved using K-means clustering, for the combination of Zero crossing Rate, Roll off, Brightness and Irregularity audio descriptors, for 3 singers for a studio recorded audio input from North Indian Classical Music recordings with the accompanying background instrument