Indian Folk Music Classification

Abstract

The goal is to classify the folk songs of India based on regions. According to the region, basically India is divided into 4 parts for the music: North, East, West and South. But for each part there are also several sub regions which differs from each-others by culture, music, speech and more. Concentrating on folk music, to take folk songs of each and every region is better than to take generalized collection of folk songs of mainly divided regions. It can give better classification of the folk songs. For this task, no any dataset is available directly. Because of that we collected songs from different websites. We could collect the folk songs of five regions: Assamese, Marathi, Kashmiri, Kannada and Uttarakhandi. After collecting folk songs, we extract various features such as Mel Frequency Cepstral Coefficients (MFCC), RMS value (loudness feature) and Spectral Centroid from each folk song for the classification of folk songs. Then we apply different classification techniques Artificial Neural Network (ANN), K Nearest Neighbor (KNN), Random Forest and Support Vector Machine (SVM) of machine learning to classify the feature sets into different classes, where each class represents a region. Subsequently, we experimented for feature subset selection methods: SelectFromModel and Recursive Feature Elimination (RFE) to get which method is appropriate for out project to get best features from the dataset to perform better classification. We collect all the results and decide which model is better for the folk song classification and how different features and combination of all features affect the classification of folk songs. Then we decide ANN performs better for our purpose than KNN and random Forest classifiers. After deciding the classifiers, we proposed a method to compare the prediction results of two finalized datasets predicted with the use of ANN by using nested K fold Cross Validation for 5 folds with the prediction of feature sets of another classifiers SVM to check and compare which dataset performs better with respect to finally received prediction result.