Beat Histogram Features from NMF-Based Novelty Functions for Music Classification

In this paper we present novel rhythm features derived from drum tracks extracted from polyphonic music and evaluate them in a genre classification task. Musical excerpts are analyzed using an optimized, partially fixed Non-Negative Matrix Factorization (NMF) method and beat histogram features are calculated on basis of the resulting activation functions for each one out of three drum tracks extracted (Hi-Hat, Snare Drum and Bass Drum). The features are evaluated on two widely used genre datasets (GTZAN and Ballroom) using standard classification methods, concerning the achieved overall classification accuracy. Furthermore, their suitability in distinguishing between rhythmically similar genres and the performance of the features resulting from individual activation functions is discussed. Results show that the presented NMF-based beat histogram features can provide comparable performance to other classification systems, while considering strictly drum patterns.