When people think about structure in music, they often think about what is written in the score or its compositional form (larger parts that conform a piece). This is important when analyzing what the composer was trying to achieve when creating a musical work. However, this is sometimes reserved for academic and theoretical research fields where knowing how to decipher a score is relevant.
Example of Score structures (taken from Allegraud, Pierre, et al. TISMIR 2019 )
Traditionally, music structure analyses have been based on structures found in music compositions. For example, Lerdahl and Jackendoff published in 1983 a central model of Music Structure from the point of view of Cognitive Science in the Generative Theory of Tonal Music (GTTM). It proposed a rule-based model that categorized musical structures' segmentation based on grouping, meter, pitch, and time-span. They would later refine the model to include notably aspects of harmonic tension in Tonal Pitch Space.
Later computational approaches like the Melisma Music Analyzer program from David Temperley were created for analyzing music structures based on the GTTM model.
Current models focus on probabilistic and machine learning approaches where researches use large amounts of data to construct models that estimate the types, location and likelihood of structural elements in a musical composition.
Citizen Science is especially relevant for the context of structure annotation, where tasks cannot be automated, and machines fail to provide good enough data. Originally, large databases containing analyses or annotations of musical pieces were impractical, given the tremendous effort and time needed for experts to manually mark each work.
Example of the SALAMI annotation interface
New technologies and the use of crowd-sourcing techniques and other Citizen Science approaches allowed the creation of new databases with MIDI and audio data and many metadata properties for research. For example, the renowned RWC and SALAMI databases contain structural annotations, and new databases such as GiantMIDI or Music4all provide score alignment, lyrics, popularity, and more. These pre-existing datasets focus on composed structures. To our knowledge, a database focusing on performed structures does not yet exist.
Another approach to the segmentation problem is to focus on how the listener's annotations are produced. Different annotations of a performance can diverge because performers and/or listeners pay attention to different musical features like tempo or timbre. After marking a performance, musical features can be computed from the signal/MIDI and then used to retrospectively reconstruct and compare the new structural boundaries created from what listeners were most likely paying attention to.
Music performance has mostly been studied in academic circles as a complement to the score, that is, the performer is a mere intermediary between the composer and the listener. However, music is a primarily auditory art form, and there many subtle, but important choices that a performer makes that tend to go unnoticed when focusing only on the score.
These decisions shape the sound that the audience will ultimately hear. Often, performers will mark notes on a score as a way of signaling what they intend to do with it.
