Meter Detection in Symbolic Music Using Inner Metric Analysis

Abstract

In this paper we present PRIMA: a new model tailored to symbolic music that detects the meter and the first downbeat position of a piece. Given onset data, the metrical structure of a piece is interpreted using the Inner Metric Analysis (IMA) model. IMA identifies the strong and weak metrical positions in a piece by performing a periodicity analysis, resulting in a weight profile for the entire piece. Next, we reduce IMA to a feature vector and model the detection of the meter and its first downbeat position probabilistically. In order to solve the meter detection problem effectively, we explore various feature selection and parameter optimisation strategies, including Genetic, Maximum Likelihood, and Expectation-Maximisation algorithms. PRIMA is evaluated on two datasets of MIDI files: a corpus of ragtime pieces, and a newly assembled pop dataset. We show that PRIMA outperforms autocorrelationbased meter detection as implemented in the MIDItoolbox on these datasets.