Cent filter-banks and its relevance to identifying the main song in carnatic music

Carnatic music is a classical music tradition from Southern India. It is primarily based on vocal music, where the lead performer is a singer. A typical Carnatic music concert is made up of several items. Each item can be made up of a number of segments, namely, monophonic vocal solo, monophonic violin solo, polyphonic (vocal and accompanying instruments) composition (or song) and monophonic percussion (thaniavarthanam). The composition (or song) segment is mandatory in every item. The identification of composition segments is necessary to determine the different items in a concert. Owing to the improvisation possibilities in a composition, the compositional segments can further consist of monophonic segments. The objective of this paper is to determine the location of song segments in a concert. The improvisational aspects of a concert lead to the number of applauses being much larger than the number of items. The concert is first segmented using the applauses. Next, inter-applause segments are classified as vocal solo, violin solo, composition and thaniavarthanam segments. Unlike Western music, the key used for different items in the concert is fixed by the performer. The key also referred to as tonic can vary from musician to musician and can also vary across concerts by the same musician. In order to classify different inter-applause segments across musicians, the features must be normalised with respect to the tonic. A new feature called Cent Filter-bank based Cepstral Coefficients (CFCC) that is tonic invariant is proposed. Song identification is performed on 50 live recordings of Car-natic music. The results are compared with that of the Mel Frequency Cepstral Coefficients (MFCC), and Chroma based Filter-bank Cepstral Coefficients (ChromaFCC). The song identification accuracy with MFCC is 80%, with CFCC features is 95% and with ChromaFCC features is 75 %. The results show that CFCC features give promising results for Carnatic music processing tasks