An Analysis of Automatically Generated Music

In recent years, there has been an explosion of research into the automatic generation of music, both audio and symbolic. Countless deep learning approaches in particular have been proposed, using a wide range of methods and producing an equally wide range of outputs. However, the evaluation of such generations is very difficult, as the gold standard method of evaluation (listening experiments with musically-trained test participants) is expensive, in terms of both time and money (assuming the participants are fairly compensated), particularly when an extensive comparative evaluation is desired. Recent work [Yi23] has undertaken such a procedure, releasing human expert ratings and generated examples comparing human compositions to automatic compositions by several methods. We take the same generations (MIDI files of classical string quartets and piano improvisations), and analyze them instead statistically, comparing properties such as rhythmic density and pitch range across each of the methods and styles. We make no claim that our analysis represents an evaluation of the selected methods, but present our findings as an exploratory look at musically-relevant statistical properties of the outputs of each method, and draw conclusions based on that.