Automatic extraction of approximate repetitions in polyphonic midi files based on perceptive criteria

In the context of musical analysis, we propose an algorithm that automatically induces patterns from polyphonies. We define patterns as "perceptible repetitions in a musical piece". The algorithm that measures the repetitions relies on some general perceptive notions: it is non-linear, non-symetric and nontransitive. The model can analyse any music of any genre that contains a beat. The analysis is performed into three stages. First, we quantize a MIDI sequence and we segment the music in "beat segments". Then, we compute a similarity matrix from the segmented sequence. The measure of similarity relies on features such as rhythm, contour and pitch intervals. Last, a bottom-up approach is proposed for extracting patterns from the similarity matrix. The algorithm was tested on several pieces of music, and some examples will be presented in this paper.