An Auditory Streaming Approach for Melody Extraction from Polyphonic Music

This paper proposes an efficient approach for the identification of the predominant voice from polyphonic musical audio. The algorithm implements an auditory streaming model which builds upon tone objects and salient pitches. The formation of voices is based on the regular update of the frequency and the magnitude of so called streaming agents, which aim at salient tones or pitches close to their preferred frequency range. Streaming agents which succeed to assemble a big magnitude start new voice objects, which in turn add adequate tones. The algorithm was evaluated as part of a melody extraction system during the MIREX audio melody extraction evaluation, where it gained very good results in the voicing detection and overall accuracy.