Blast for Audio Sequences Alignment: a Fast Scalable Cover Identification

Benjamin Martin (1) , D.G. Brown (2) , Pierre Hanna (1) , Pascal Ferraro (3)

1 Image et Son
2 David R. Cheriton School of Computer Science
3 LaBRI - Laboratoire Bordelais de Recherche en Informatique

Searching for similarities in largemusical databases is common for applications such as cover song identification. These methods typically use dynamic programming to align the shared musical motifs between subparts of two recordings. Such music local alignment methods are slow, as are the bioinformatics algorithms they are closely related to. We have adapted the ideas of the Basic Local Alignment Search Tool (BLAST) for biosequence alignment to the domain of aligning sequences of chroma features. Our tool allows local music sequence alignment in near-linear time. It identifies small regions of exact match between sequences, called seeds, and builds local alignments that include these seeds. Seed determination is a key issue for the accuracy of the method and closely depends on the database, the representation and the application. We introduce a particular seeding approach for cover detection, and evaluate it on both a 2000-piece training set and themillion song dataset (MSD). We show that the heuristic alignment drastically improves time computation for cover song detection. Alignment sensitivity is still very high on the small database, but is dramatically weakened on the MSD, due to differences in chroma features. We discuss the impact of different choices of these features on alignment of musical pieces.

Domaines: Son [cs.SD]

Benjamin Martin, D.G. Brown, Pierre Hanna, Pascal Ferraro. Blast for Audio Sequences Alignment: a Fast Scalable Cover Identification. 13th International Society for Music Information Retrieval Conference, 2012, Portugal. p. 529-534. ⟨hal-01006094⟩