Audio similarity matrices enhancement in an image processing framework

Audio similarity matrices have become a popular tool in the MIR community for their ability to reveal segments of high acoustical self-similarity and repetitive patterns. This is particularly useful for the task of music structure segmentation. The performance of such systems however relies on the nature of the studied music pieces and it is often assumed that harmonic and timbre variations remain low within musical sections. While this condition is rarely fulfilled, similarity matrices are often too complex and structural information can hardly be extracted. In this paper we propose an image-oriented pre-processing of similarity matrices to highlight the conveyed musical information and reduce their complexity. The image segmentation processing step handles the image characteristics in order to provide us meaningful spatial segments and enhance thus the music segmentation. Evaluation of a reference structure segmentation algorithm using the enhanced matrices is provided, and we show that our method strongly improves the segmentation performances.