A methodology for procedural piano music composition with mood templates using genetic algorithms

Abstract

Creating music in an automatic way has been studied since the beginning of artificial intelligence. One of the biggest obstacles of music generation is the vagueness and subjectivity of the mood or emotion transmitted by a music piece. In this work, we experiment with the generation of piano music using template pieces, represented in MIDI format, as a mood directive. We generated a population of random pieces for templates of two opposing moods - happy and sad - and evolved them with a genetic algorithm until their intended mood was close enough to their respective templates. The fitness function that we implemented uses MIDI statistical features to calculate the distance between the given piece and the template. The generated music pieces were evaluated by human listeners thorough a questionnaire. This evaluation has shown that the generated music pieces were able to express the same mood as the template. However, they still sounded computer-generated, probably due to the lack of rhythm regularity and synchronicity.