A Case Study Towards Segment-Aware Version Identification

Version identification (VI) or cover song identification is the task of automatically detecting whether musical tracks are originating from the same work. An important step in approaches for solving this task is the shingling along the time axis. Recently proposed models show a better retrieval accuracy when using shorter shingles (e.g., of 20 seconds) rather than relying on longer ones (e.g., more than 1 minute) or even full tracks. However, all current approaches define a fixed length for the shingles, even though the actual segments in a musical sense (e.g., chorus) are usually varying in length and might even vary between different versions of the same musical work. This case study explores new perspectives on VI beyond fixed-length shingles. Based on a new VI dataset with manually annotated segment labels, we investigate the distributions of pairwise distances of version embeddings obtained from a state-of-the-art VI model. We further examine the impact of different shingle-to-segment offsets to uncover the potential performance degradation in current VI testing methods.