Makam recognition using extended pitch distribution features and multi-layer perceptrons

Abstract:

This work focuses on automatic makam recognition task for Turkish makam music (TMM) using pitch distributions that are widely used in mode recognition tasks for various music traditions. Here, we aim to improve the performance of previous works by extending distribution features and performing parameter optimization for the classifier. Most music theory resources specifically highlight two aspects of the TMM makam concept: use of microtonal intervals and an overall melodic direction that refers to the design of melodic contour on the song/musical piece level. Previous studies for makam recognition task already utilize the microtonal aspect via making use of high resolution histograms (using much finer bin width than one 12th of an octave). This work considers extending the distribution feature by including distributions of different portions of a performance to reflect the long-term characteristics referred in theory for melodic contour, more specifically for introduction and finalis. Our design involves a Multi-Layer Perceptron classifier using an input feature vector composed of pitch distributions of the first and the last sections together with the overall distribution, and the mean accuracy of 10 iterations is 0.756. The resources used in this work are shared for facilitating further research in this direction.