Generative structures in improvisation : computational segmentation of keyboard performances

We assessed the hypothesis elaborated by Pressing and by the first author that improvisation is normally based on the generation of segmental contrasts in musical features. Nine experienced improvisers performed a series of free- and 3-section referent-based pieces, each of about three minutes. Each player undertook eight improvisations with 'referent' instructions, preceded and followed by a free improvisation. A MIDI-equipped grand piano was used, and audio and MIDI were recorded. Computational analyses of the MIDI data assessed whether performers realized the referents successfully, and then determined the large-scale segmentation of the free improvisations. The performers almost always fulfilled the referents (68/72 improvisations, p < 0.0001), and their free improvisations were also susceptible to large-scale segmentation (p < 0.005). Since performers were only told of the referent structures as the session proceeded, the similar segmentation of both the first and second free improvisation suggests that such segmentation is common in solo free improvisation, in accord with the hypothesis.