PAUL-2: An Upgraded Transformer-Based Redesign of the Algorithmic Composer PAUL

Abstract

In this paper, we introduce $\mathsf {PAUL \text {-}2}$, an algorithmic composer for two-track piano pieces of specifiable difficulty levels, as a ground-up redesign of its predecessor system $\textsf{PAUL}$. While $\textsf{PAUL}$ was designed using a long short-term memory neural network, along with a sequence-to-sequence network, $\mathsf {PAUL \text {-}2}$ is based on the state-of-the-art transformer architecture and makes use of relative attention. A shortcoming of $\textsf{PAUL}$ was that it generated unsatisfying accompanying tracks and allowed for only few difficulty levels. $\mathsf {PAUL \text {-}2}$ overcomes these limitations and theoretically supports an arbitrary number of difficulty classes due to the fact that it utilises an additional encoder for handling difficulty information. We also carried out a medium-scale survey which showed that the output of PAUL-2 was evaluated quite favourably by the participants.

We would like to thank Wolfgang Schmidtmayr from the University of Music and Performing Arts Vienna and Geraldine Fitzpatrick from the Human-Computer Interaction Group at our university for their valuable input. We also would like to thank the anonymous reviewers for their helpful comments.