Abstract
Partitioning a recording into non-overlapping time intervals comes in many forms. There is the structural segmentation task which labels structures either syntactically as A and B, or structurally as verse or chorus. The chord annotation task is similar, labeling segments by their chords. While many of these annotations are flat, this article extends the method by McFee and Kinnaird (2019) for automatically enhancing structural annotations by inferring (and expanding) hierarchical information from the segment labels. One of our extensions adds new rules that allow for structural labels with a wider vocabulary than the syntactical ones in the SALAMI dataset. Using this first extension, we compare annotations from the Beatles-TUT and Isophonics datasets to investigate similarities between these annotations. Our second extension creates a multi-level annotation for chords that addresses a number of current challenges in chord evaluation. Using a large collection of chord annotations (manually and automatically generated), we investigate how and where the multi-level hierarchies can enhance (or detract from) comparing chord annotations.