Trading Fours with Mr. PC: Toward an Improved Model for Jazz Synthesis

Machine learning (ML) is a demonstrably effective method for learning pattern- and rule-based systems, such as natural language or music. This thesis identifies jazz improvisation as one such system and seeks to design an ML model for learning it. To this end, Mr. PC is an ML model capable of synthesizing novel jazz improvisation in a swing style. It was constructed using a state-of-the-art Transformer-XL architecture and trained with the hereunto unexplored Filosax dataset. The Transformer-XL architecture is a promising solution for inducing long-term structure in generated improvisation: a persistent challenge in the music generation literature. The quality of the generated improvisation is evaluated using both traditional music analysis techniques and quantitative fractal analysis. Further, the output is compared to that of a smaller, simpler model trained on the same data set to test the margin of improvement.