Raag Recognition Using Pitch-Class and Pitch-Class Dyad Distributions.
Parag Chordia, Alex Rae · International Symposium/Conference on Music Information Retrieval · 2007
We describe the results of the first large-scale raag recognition experiment. Raags are the central structure of Indian classical music, each consisting of a unique set of complex melodic gestures. We construct a system to recognize raags based on pitch-class distributions (PCDs) and pitch-class dyad distributions (PCDDs) calculated directly from the audio signal. A large, diverse database consisting of 20 hours of recorded performances in 31 different raags by 19 different performers was assembled to train and test the system. Classification was performed using support vector machines, maximum a posteriori (MAP) rule using a multivariate likelihood model (MVN), and Random Forests. When classification was done on 60s segments, a maximum classification accuracy of 99.0% was attained in a cross-validation experiment. In a more difficult unseen generalization experiment, accuracy was 75%. The current work clearly demonstrates the effectiveness of PCDs and PCDDs in discriminating raags, even when musical differences are subtle.