Proceedings of the 20th ISMIR Conference, Delft, Netherlands
Structural segmentation is the task of partitioning a recording into non-overlapping time intervals, and labeling each segment with an identifying marker such as A, B, or verse. Hierarchical structure annotation expands this idea to allow an annotator to segment a song with multiple levels of granularity. While there has been recent progress in developing evaluation criteria for comparing two hierarchical annotations of the same recording, the existing methods have known deficiencies when dealing with inexact label matchings and sequential label repetition. In this article, we investigate methods for automatically enhancing structural annotations by inferring (and expanding) hierarchical information from the segment labels. The proposed method complements existing techniques for comparing hierarchical structural annotations by coarsening or refining labels with variation markers to either collapse similarly labeled segments together, or separate identically labeled segments from each other. Using the multi-level structure annotations provided in the SALAMI dataset, we demonstrate that automatic hierarchy expansion allows structure comparison methods to more accurately assess similarity between annotations.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
© Brian McFee, Katherine M. Kinnaird. Licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0).
Brian McFee, Katherine M. Kinnaird. “Improving structure evaluation through automatic hierarchy expansion”, 20th International Society for Music Information Retrieval Conference, Delft, The Netherlands, 2019.