Improving algorithmic music composition with machine learning

Algorithmic generation of musical sounding music is an interesting but challenging task, because machines do not inherently possess any form of creativity, which is necessary to create music. The Automated Composer of Style Sensitive Music II (ACSSM II) system generates music by searching for a sequence of music segments that best satisfy various constraints, including length and pitch range, harmonic backbone, and consistency with a probabilistic model of a composer’s style. As with all optimization problems, our problem requires the construction of a search space; we utilize a clustering space produced by grouping together music segments having similar musical features. The output sequence is simply a path passing through these clusters. In order to produce such a sequence, we utilize a genetic algorithm. To evaluate the system, we have conducted an experiment, involving five subjects who possess at least three years of musical training, in which the overall musical quality of produced music was assessed. Our results show that the automatically generated music achieved a mean satisfaction score of 7.5/10, which is significantly higher than that given to the music produced by the earlier ACSSM system. Hence, the results suggest that ACSSM II is a better system than its predecessor and is capable of generating reasonably musical sounding music.