Musical Similarity in a Polyphonic Context : a Model outside Time

In the context of pattern extraction from polyphonic music, we challenge an approach outside time for computing the similarity between two musical sequences which neither modelizes temporal context nor expectancy. If theses notions might play a role in our perception of musical patterns, we propose in a first step to investigate the limits of a system that ignores them. Our approach relies on a new representation of the polyphonic musical sequence which is quantized in equally-spaced beat-segments and on a new definition of the notion of similarity in a polyphonic context. In agreement with ([1], [2]), we think that text-matching methods, or pure mathematical algorithms are not directly convenient for music analysis. We think that the similarity relationships between musical sequences are the result of a cognitive process that implies to evaluate the algorithms in terms of their cognitive relevance. As few experiments have been made on people's cognitive criteria for similarity measuring, we base our criteria on heuristics that were inspired from some musical issues. Three different sets of features have been considered: pitches, pitch contours and rhythm. For each set, a similarity measure is computed. The global similarity value results from the linear combination of the three values. The algorithm was tested on several pieces of music, and interesting results were found. At the same time, new questions were raised on the notion of similarity (this research is part of the European project Cuidado).