DETECTION OF ORCHESTRAL BLENDS IN MUSICAL SCORES

In music orchestration, orchestral blend refers to the process in which multiple instruments, when played together under certain conditions, merge their individual sounds to form a cohesive and distinct timbre. Despite being primarily a perceptual phenomenon, previous studies have shown that some compositional aspects of music, such as the synchronization between different instrumental parts, can promote the generation of orchestral blends. Those characteristics can be exploited by composers who want to write blends. We hypothesize therefore that blends can be detected starting from the score only. In this paper, we propose a machine learning approach with perceptioninformed features to identify blending instruments in orchestral scores. We propose and use several evaluation metrics and show that our model improves the state-ofthe- art algorithm. Moreover, we show that a simple algorithm based on comparing the number of notes played concurrently by two instruments is sufficient to obtain better results on these metrics than the state-of-the-art algorithm. This study confirms with a score-based approach the impact of music cues such as synchrony, harmonicity, parallelism, and other descriptors of the similarity between instrumental parts on the perception of orchestral blends.