A Simple, High-Yield Method for Assessing Structural Novelity

Abstract

The structural dimension of music plays an important role in its affective appreciation. One particular aspect is related to the temporal succession of moments, each characterized by particular musical properties. One classical approach in computational modelling of this aspect is based on similarity matrix representations, where successive states are visualized by successive squares along the main diagonal, bearing some resemblance to checkerboards. One referential method estimates a so-called novelty curve, representing the probability along time of the presence of transitions between successive states, as well as their relative importance. Novelty is traditionally computed by comparing – through cross-correlation – local configurations along the diagonal with an ideal checkerboard kernel. The method is limited by a strong dependency on kernel size, which imposes a single level of granularity in the analysis and fails to grasp common musical structures made of a succession of states of various sizes. We introduce a simpler but more powerful and general method that automatically detects homogeneous segments of any size. Only half of the similarity matrix is retained, in order to compare each new instant solely with the past and exclude the future. For each instant in the piece, novelty is assessed by first determining the temporal scale of the preceding homogeneous part as well as the degree of contrast between that previous part and what just comes next. Detailed results show how and why this method offers a richer and more intuitive structural representation encompassing all granularity levels.