Variation Transformer: New datasets, models, and comparative evaluation for symbolic music variation generation

        Variation in music is defined as repetition of a theme, but with various modifications, which plays an important role in many musical genres in developing core music ideas into longer passages. Existing research on variation in music is mostly confined to datasets consisting of classical theme-and-variation pieces, and generative models limited to melody-only representations. In this paper, to address the problem of the lack of datasets, we propose an algorithm to extract theme-variation pairs automatically, and use it to annotate two called POP909-TVar (2,871 theme-variation pairs) and VGMIDI-TVar datasets (7,830 theme-variation pairs). We propose both non-deep learning and deep learning based symbolic music variation generation algorithms, and report the results of a listening study and featured-based evaluation for these models. One of our two newly proposed models, called Variation Transformer, outperforms all other models that listeners evaluated for "variation success", including non-deep learning and deep learning based approaches. An implication of this work for the wider field of music making is that we now have a model that can generate relatively novel material, but with stronger and perceivably more successful relationships to some given prompt or theme.

Variation samples generated by models trained on the POP909-TVar dataset

These 15 groups of stimuli were picked randomly from POP909-TVar generated outputs, and used in our listen study. Each row below shows a theme and its variations, where the variations are either composed by human composers or generated by algorithms.

Variation samples generated by models trained on the VGMIDI-TVar dataset

These 15 groups of stimuli were picked randomly from VGMIDI-TVar generated outputs, and used in our listen study. Each row below shows a theme and its variations, where the variations are either composed by human composers or generated by algorithms.