Sampling Variations of Sequences for Structured Music Generation

Recently, machine-learning techniques have been successfully used for the generation of complex artifacts such as music or text. However, these techniques are still unable to capture and generate artifacts that are convincingly structured. In particular, musical sequences do not exhibit pattern structure, as typically found in human composed music. We present an approach to generate structured sequences, based on a mechanism for sampling efficiently variations of musical sequences. Given an input sequence and a statistical model, this mechanism uses belief propagation to sample a set of sequences whose distance to the input sequence is approximately within specified bounds. This mechanism uses local fields to bias the generation. We show experimentally that sampled sequences are indeed closely correlated to the standard musical similarity function defined by Mongeau and Sankoff. We then show how this mechanism can be used to implement composition strategies that enforce arbitrary structure on a musical lead sheet generation problem. We illustrate our approach with a convincingly structured generated lead sheet in the style of the Beatles.