Determination of the meter of musical audio signals: Seeking recurrences in beat segment descriptors

We address the problem of classifying polyphonic musical audio signals by their meter: the number of beats between regularly recurring accents (or downbeats). The problem is simplified to a ‘duple ’/‘triple’ decision. Experiments have been conducted on a 70 instances database (20s excerpts from pieces of music without particular genre nor timbre restriction). Our approach aims to test the hypothesis that acoustic evidences for downbeats can be measured on signal low-level features; focusing especially on their temporal recurrences. We experimented several approaches to the problem of feature selection and report some interesting results: measurements of a very small set of beat descriptors (i.e. 4) and subsequent processing (based on autocorrelation functions) permit to reach around 95% of correct classification. Using only the temporal centroid, almost 90% of correct classification can be achieved.