Melody Extraction from MIDI Files Using AI-ML Techniques

Development of Artificial Intelligence and Machine Learning (AI-ML) based automated tools for melody extraction from audio music files are severely handicapped due to the scarcity of proper musical annotation to act as labels for supervised learning. For many kinds of traditional music the melodies can be quite intricate, and detailed annotation is not only scarce but also quite difficult to create. MIDI music files on the other hand carry within them the notes of melody which if extracted can act as musical annotations for training supervised melody extraction tools. In this study, several Machine Learning based classifiers were trained to determine whether a particular channel of MIDI file carries a melody or not. The six features used for training were Average TFIDF, Overlap, Number of Unique Notes, Note Duration, Note Velocity and Entropy. The testing accuracies obtained by Support Vector Machine (SVM), Random Forest, Logistic Regression, and Artificial Neural Network (ANN) classifiers were 86.67%, 88.64%, 86.91% and 93.80% respectively, with ANN based method giving the best testing result.