Musical Structure Retrieval by Aligning Self-Similarity Matrices

We propose a new retrieval system based on musical structure using symbolic structural queries. The aim is to compare musical form in audio files without extracting explicitly the underlying audio structure. From a given or arbitrary segmentation, an audio file is segmented. Irrespective of the audio feature choice, we then compute a selfsimilarity matrix whose coefficients correspond to the estimation of the similarity between entire parts, obtained by local alignment. Finally, we compute a binary matrix from the symbolic structural query and compare it to the audio segmented matrix, which provides a structural similarity score. We perform experiments using large databases of audio files, and prove robustness to possible imprecisions in the structural query.