Bridging Subjectivity and Objectivity in Evaluation of Machine-Generated Jazz Music: A Multimetric Approach

This research explores different musical metrics to bridge the gap between subjectivity and objectivity in jazz music generation by employing diverse metrics including pitch class histogram entropy, groove pattern similarity, and creating two new metrics based on jazz theoretical concepts: the chord tone emphasis and the swing deviation metric. The study utilizes both a Markov model and a Recurrent Neural Network for music generation, with a dataset consisting of choruses of F-blues. A human survey is conducted on the preferred model for jazz generation to align these objective metrics with the subjective preference of humans. The results suggest that combining these objective metrics with subjective survey responses provides a more comprehensive understanding of generative jazz music’s quality. While challenges remain, these metrics contribute to a more nuanced and balanced evaluation of the artistic output. Future work may involve exploring the integration of additional metrics and theoretical concepts to further enrich the evaluation process.