Algorithmic Music Composition Using Probabilistic Graphical Models and Artificial Neural Networks

Composing music algorithmically has been a goal long-pursued by many computer scientists. Various methods have been implemented to achieve this, ranging from randomly selecting musical components to deep learning models. The main focus of this research is to develop a model which fools a human into believing the output music is human-made. This study uses a collection of rock music MIDI files which the notes, chords, pitches and duration are extracted as features. A Bayesian network is selected as the main model for this research and a Long Short-Term Memory (LSTM) network as the benchmark model. A Turing test was performed on 20 people for both models and the LSTM on average was identified as human-made 36% of the time, while the Bayesian network, on average, had been misidentified 39% of the time. These results may indicate that music is more probabilistic than time-dependent.