Generating Subjects for Pieces in the Style of Bach’s Two-Part Inventions

We describe how to begin the computational process of composing a piece in the style of J.S. Bach’s two-part inventions by automatically generating plausible musical subjects. A generate and test protocol is proposed, whereby subjects would be generated by a modified random sampling technique and then tested against information theoretic measures, with the aim of filtering out unsuitable subjects. The statistical models to be sampled were constructed by machine learning from a corpus comprising the first few bars of the upper parts of the fifteen two-part inventions, using multiple viewpoint systems. Using information content, we were able to correctly classify subjects as suitable or unsuitable from a pitch structure point of view in 115 out of 120 cases. Mean information content based on a simple short-term model (i.e., the statistics were taken from the generated subject, not from the corpus) was partially successful in filtering out unduly repetitive subjects