The Chordinator: Modeling Music Harmony by Implementing Transformer Networks and Token Strategies

Abstract

This paper compares two tokenization strategies for modeling chord progressions using the encoder transformer architecture trained with a large dataset of chord progressions in a variety of styles. The first strategy includes a tokenization method treating all different chords as unique elements, which results in a vocabulary of 5202 independent tokens. The second strategy expresses the chords as a dynamic tuple describing root, nature (e.g., major, minor, diminished, etc.), and extensions (e.g., additions or alterations), producing a specific vocabulary of 59 tokens related to chords and 75 tokens for style, bars, form, and format. In the second approach, MIDI embeddings are added into the positional embedding layer of the transformer architecture, with an array of eight values related to the notes forming the chords. We propose a trigram analysis addition to the dataset to compare the generated chord progressions with the training dataset, which reveals common progressions and the extent to which a sequence is duplicated. We analyze progressions generated by the models comparing HITS@k metrics and human evaluation of 10 participants, rating the plausibility of the progressions as potential music compositions from a musical perspective. The second model reported lower validation loss, better metrics, and more musical consistency in the suggested progressions.

This paper is an outcome of: MUSAiC, a project that has received funding from the European Research Council under the European Union’s Horizon 2020 research and innovation program (Grant agreement No. 864189); and a Margarita Salas Grant, UPF, Barcelona, Spain.