A Music-Theoretically Informed Metric for Evaluating Generated Harmonic Progressions

Lorenz Zauter

Abstract

The coherence of a musical piece is largely determined by its phrases, which in classical music are characterized by the melody, by voice leading conventions and harmonic rules. While these rules can be formalized, composers often bend them, adding to the music’s appeal. Ensuring coherence without monotony in music generation is crucial. Measuring the completeness of a musical phrase in a data-driven way, adaptable to different composers and corpora, has not been done until now. This paper proposes a metric to evaluate harmonic progression completeness by learning syntactic closure from the data rather than presupposing grammar. We embed chords similarly to word embeddings in Natural Language Processing, cluster chords with similar harmonic functions, and factor progressions into their ‘core’ and cycles of repeating chords. This method calculates the probability of a sequence being complete and assesses adherence to the musical style of the corpus. Existing methods focus on local prediction accuracies and chord similarities, but lack a data-driven metric for evaluating progression completeness. This paper lays the groundwork for such a metric in western classical music.