A bi-stage approach to North Indian raga distinction

Music embraces an intricate assembly of auditory elements, carefully organized in diverse configurations to articulate a variety of human emotions, moods, thoughts, feelings, temporal contexts, and situations. One of the primary aspects of any music composition is the raga which governs its melodic framework. Thus, the delineation of ragas assumes an enormous significance as a preliminary step preceding a more deeper and intricate analysis. Each of these ragas is meant to be practiced during a particular time of the day to amplify the emotional content and physical involvement. In this current work, a machine learning-based approach has been proposed to classify the dawn and dusk time (Sandhi Prakash) ragas. Here, mel-frequency cepstral coefficients (MFCC) based feature extraction technique has been applied which was further processed to generate second-level statistical features. This brought down the original feature dimension by means of effective representation of the raw features. Several classification techniques were employed and a new bi-stage raga distinction technique has been proposed. The first stage classifies ragas as dawn/ dusk while the second stage performs deeper classification for these groups separately to identify the exact raga. Experiments were performed with over 57K clips from 11 ragas belonging to the 2 time periods and a performance improvement of 0.7% was obtained for the dusk ragas using the bi-stage approach over the single shot classification technique. The highest possible accuracy of 96.47% was obtained for distinguishing the dusk ragas with only 2-second clips in the experiments.