A Novel Method to Identify Audio Descriptors, Useful in Gender Identification from North Indian Classical Music Vocal

Gender identification application of Music Information Retrieval (MIR) research requires that the audio features that are extracted from the sound files should be sufficiently strong to be able to correctly classify the singer into male or a female. In literature, a lot of such applications are developed for speaker’s gender identification. Unfortunately, these applications cannot be applied for gender identification of a singer. For singer identification, concept of sound timbre along with interference of background music has to be considered. Merely applying filters to this background voice may help for polyphonic music recordings, such as, western classical or popular music, but in case of North Indian classical music, which is homophonic, the continuous interference of a supportive musical instrument such as ‘Tanpura’ or ‘Harmonium’ cannot be neglected or filtered out. In this paper, we have introduced a novel indirect method to identify the audio descriptors that are helpful in identifying the gender of a singer singing north Indian classical music. The aim is to run a two folded algorithm in which, first, we identify the singer with the help of various combinations of traditional audio descriptors under timbre taxonomy and then to backtrack and look back towards the audio descriptors that made this successful identification of the singer. A traditional K-means classifier is used. The algorithm made run on two databases, viz. DB1 containing all male singers and DB2 containing all female singers. The results showed that, for male singer along with Zero crossing rate and MFCC (Mel Frequency Cepstrum Coefficients), there are two more audio descriptors that take part in achieving highest efficiency of singer identification. These are Roughness and Irregularity. While for female these were Roll off and Brightness along with Zero crossing Rate and MFCC. The accuracy of singer identification for male and female singers is found to be 83.33% and 85% respectively.