Recognition of the Main Melody in a Polyphonic Symbolic Score using Perceptual Knowledge

It is in many cases easy for a human to identify the main melodic theme when listening to a music example. Melodic properties have been studied in several research projects, however, the differences between properties of the melody and properties of the accompaniment (non-melodic) voices have not been addressed until recently. A set of features relating to basic low-level statistical measures were selected considering general perceptual aspects. A new 'narrative' measure was designed intended to capture the amount of new unique material in each voice. The features were applied to a set of scores consisting of about 250 polyphonic ringtones consisting of MIDI versions of contemporary pop songs. All tracks were annotated into categories such as melody and accompaniment. Both multiple regression and support vector machines were applied on either the features directly or on a Gaussian transformation of the features. The resulting models predicted the correct melody in about 90% of the cases using a set of eight features. The results emphasize context as an important factor for determining the main melody. A previous version of the system has been used in a commercial system for modifying ring tones.