Automatic Raag Classification of Pitch-tracked Performances Using Pitch-class and Pitch-class Dyad Distributions
Parag Chordia · University of Michigan Library Repository · 2006
A system was constructed to automatically identify raags using pitch-class (PCDs) and pitch-class dyad distributions (PCDDs) derived from pitch-tracked performances. Classification performance was 94% in a 10-fold cross-validation test with 17 target raags. Using PCDs alone, performance was 75%, and 82% using only PCDDs. Best performance was attained using a maximum a posteriori (MAP) rule with a multivariate normal (MVN) likelihood model. Each raag was divided into non-overlapping 30 second segments and pitch tracked using the Harmonic Product Spectrum (HPS) algorithm. Pitch tracks were then transformed into pitch-class sequences by segmenting into notes using a complex-domain detection function. For each note, pitch-class was determined by taking the mode of the detected pitches from the onset of the note to the next onset. For the given tuning, the nearest pitch was found based on a just-intoned chromatic scale. The comparison was made in the log-frequency domain. PCDs and PCDDs were estimated from each segment leading to 12 PCD features and 144 PCDD features. Thus, each segment was represented by a 156-dimensional feature vector, representing the relative frequency of pitch-classes and pitch dyads. It was found that performance improved significantly (+15%) when principal component analysis was used to reduce the feature vector dimension to 50. The study suggests that PCDs and PCDDs may be effective features for raag classification. However, the database size must be expanded in size and diversity to confirm this more generally.