Statistical characterisation of melodic pitch contours and its application for melody extraction

In this paper we present a method for the statistical characterisation of melodic pitch contours, and apply it to automatic melody extraction from polyphonic music signals. Within the context of melody extraction, pitch contours represent time and frequency continuous sequences of pitch candidates out of which the melody must be selected. In previous studies we presented a melody extraction algorithm in which contour features are used in a heuristic manner to filter out non-melodic contours. In our current work, we present a method for the statistical modelling of these features, and propose an algorithm for melody extraction based on the obtained model. The algorithm exploits the learned model to compute a "melodiness" index for each pitch contour, which is then used to select the melody out of all pitch contours generated for an excerpt of polyphonic music. The proposed approach has the advantage that new contour features can be easily incorporated into the model without the need to manually devise rules to address each feature individually. The method is evaluated in the context of melody extraction and obtains promising results, performing comparably to a state-of-the-art heuristic-based algorithm.