Optimal filtering of dynamics in short-time features for music organization

There is an increasing interest in customizable methods for organizing music collections. Relevant music characterization can be obtained from short-time features, but it is not obvious how to combine them to get useful information. In this work, a novel method, denoted as the Positive Constrained Orthonormalized Partial Least Squares (POPLS), is proposed.

Working on the periodograms of MFCCs time series, this supervised method finds optimal filters which pick up the most discriminative temporal information for any music organization task. Two examples are presented in the paper, the first being a simple proof-of-concept, where an altosax with and without vibrato is modelled.

A more complex \$11\$ music genre classification setup is also investigated to illustrate the robustness and validity of the proposed method on larger datasets. Both experiments showed the good properties of our method, as well as superior performance when compared to a fixed filter bank approach suggested previously in the MIR literature.

We think that the proposed method is a natural step towards a customized MIR application that generalizes well to a wide range of different music organization tasks.