Algorithmic Music Composition Comparison

Abstract

This paper presents the application of Machine Learning (ML) algorithm as an algorithmic music composer, compared to a rule-based algorithm. The ML model is based on LSTMs which takes in previous notes and predicts the next set of notes based on a midi format. For the rule-based method, we apply chord progression rules and binary rhythm pattern theory. We used both algorithms to generate music in two different genres, namely rock, and jazz. To evaluate the effectiveness of the algorithms, fifteen raters are asked to identify the genre of the generated songs. The results showed 77.33% of the rulebased algorithms Jazz songs were correctly identified, compared to the 62.67% generated by the LSTM. For the rock genre, only 49.33% percent of rule-based algorithms songs and 44% Machine Learning algorithms songs were correctly identified. In terms of music satisfaction, the rule-based algorithm on average obtains higher scores in both genres, 2.17 for Jazz and 2.42 for Rock while Machine Learning algorithm receives 1.83 for Jazz songs and 1.57 for Rock.