In Indian music, including Indian classical music, one can keep the tonic at any pitch as the reference point from which other notes are realized. This is in contrast to most Western classical forms, where a particular musical piece has a specific key that determines the harmony and counterpoint. Indian music is monophonic and does not support harmony and counterpoint. But this liberty to keep the “Sa” (tonic) anywhere also makes its detection a nice research problem. The authors have rightfully targeted the joint detection of tonic and ragas in the present work.
The authors use the concept of pitch class distribution, in which musicians never use fixed pitch, the mood changes according to the pitch contour, and even the pitch movements between the notes are crucial. The authors obtain the best results with a kernel-density pitch distribution along with a nearest neighbor classifier through a Bhattacharya distance.
A surprising finding of the paper is that although a raga is a melodic structure, good classification can be done without sequential information. This is surprising, because the aroh-awaroh (ascent-descent) of notes is rather strictly sequential in a raga. The authors propose to experiment with hidden Markov models in future.
The definition of raga, given on page 83, misses the fact that a particular raga characterizes a particular mood or musical emotion. This means that while two ragas such as Bageshree and Shivranjani (which incidentally both belong to the Kafi thaat) can both evoke sadness or karuna rasa, the sadness of Bageshree is different from the sadness of Shivranjani, just as the tastes of mango and apple are different even though both may be classified as sweet. The raga rules, such as the fixing of notes, the note combinations, and the aroh-awaroh, actually build this very diversity, which is, in my opinion, more important than any commonality classification.