Classifying Carnatic Ragams Based On Audio Features

This study developed a Random Forest Classifier to identify ragams in Carnatic music by analyzing key audio features such as Chroma STFT, spectral properties, and MFCC coefficients. Trained on a dataset of 1,112 audio clips across 32 ragams, the model achieved a high classification accuracy of 99.1%, outperforming other baseline models.

Carnatic music is a form of classical singing originating from South India. It consists of many nuances and patterns that are organized into various ragams, which are melodic frameworks for a song that dictate the types of notes and modulations present. Each ragam contains complicated variations of notes and features, however, so it is often difficult for singers and listeners to identify ragams. To address this challenge, I built a model using a Random Forest Classifier that analyzes features such as Chroma Short-Time Fourier Transform, Spectral Centroid, Spectral Bandwidth, various MFCC coefficients, and the ragam label in order to identify the ragam of the audio clip. The model was trained on a dataset of 1112 audio clips that includes 32 ragams– 7 melakarta “parent” ragams and 25 janya “child” ragams (Sanjana Satish68l). Each ragam has certain patterns within its features that make it easy to distinguish from other ragams, and therefore easy for the model to categorize. By analyzing these distinct features, the model was able to accurately classify the ragam of almost all the test clips. The accuracy was considerably high: around 99.1%, which outperformed both the K Nearest Neighbors and Decision Tree models.

Explore More!

Research Paper