Machine learning system for estimating the rhythmic salience of sounds

Abstract

This article describes experimental work carried out in attempt to improve the effectiveness of musical rhythm retrieval systems. The authors define basic notions in the area of hierarchical rhythm retrieval and describe a procedure for inducing rhythmic hypotheses in a given melody. Utilizing an approach commonly used in the data mining domain, an association rule model has been applied to estimate the rhythmic salience of sounds based on the physical attributes of duration, frequency and amplitude. On the basis of the knowledge obtained by the machine learning system, the authors propose five functions to rank sounds according to their tendency to be located in accented positions in a melody. Adapted precision and recall measures were used to validate the proposed functions and conduct experimental verification. Conclusions derived from the results of the experiments have also been presented.