Algorithmic music composition involves the use of rules to generate melodies. One simple but interesting technique is to select notes sequentially according to a transition table that specifies the probability of the next note as a function of the previous context.
I describe an extension of this transition table approach using a recurrent connectionist network called CONCERT. CONCERT is trained on a set of melodies written in a certain style and then enables the composition of new melodies in the same style. A central ingredient of CONCERT is the incorporation of a psychologically-grounded representation of pitch.
CONCERT was tested on sets of examples artificially generated according to simpler rules and was shown to learn the underlying structure, even where other approaches failed. In a large experiment, CONCERT was trained on a set of J. S. Bach minuets and marches and was then allowed to compose novel melodies.
Although the compositions are pleasant, I don’t foresee a Grammy in the near future. The main problem is a lack of global coherence. Some ideas are presented about how a network can be made to induce structure at both local and global scales.