Carnatic music (CM) is a form of South Indian music characterized by continuous pitch variation called gamakas. A svara in CM is a musical note rendered with or without gamaka. Svaras are strung together to form phrases, and phrases constitute ragas. CM in general, and gamaka in particular, is taught by example (orally or on instruments). Consequently, as is typical of oral traditions, there is variation in the detail and style between different schools of CM, musicians, and even renditions. Nevertheless, the identity of a raga is preserved in its renditions, perhaps due to a common, underlying skeletal grammar. The focus of this research is a step towards understanding this grammar, namely automatic transcription of CM.

In Seminar I, we summarized the existing svara-based notation schemes and recent descriptive transcription representations. We then proposed separating CM pitch curves into constant-pitch notes (CPNs) and stationary points (STAs). In this seminar, we propose the use of the direction of the pitch curve to further separate CPNs and STAs. We define upward anchors as CPNs that are adjacent to STAs higher in pitch than the CPN, and define downward anchors similarly. Similarly, we define min-STAs and max-STAs as STAs that are local minima or maxima, respectively. For each raga, peaks detected from the histograms of upward and downward anchors, and of min-STAs and max-STAs form target notes. The targets provide information on the raga grammar and serve as the reference for our proposed automated measurement of precision in CM. The results show that the precision of CPNs and of STAs differ significantly.

We then propose state-based transcription, where we model the CPNs as being in anchor or CPN states, and STAs as being in both anchor and transient states. We use the Viterbi algorithm to trace back the most likely sequence of states. The pitch values at the STAs and CPNs are quantized to the corresponding anchor-targets or STA-targets. We conducted a listening test whose results show that the synthesis of the interpolated, quantized pitches scores significantly better than the other, most recent transcription technique. Further, unlike other transcription techniques, the state information of STAs enables slowing down renditions for teaching purposes without compromising raga characteristics.

All are cordially invited.