On the Detection of Melody Notes in Polyphonic Audio

This paper describes a method for melody detection in polyphonic musical signals. Our approach starts by obtaining a set of pitch candidates for each time frame, with recourse to an auditory model. Trajectories of the most salient pitches are then constructed. Next, note candidates are obtained by trajectory segmentation (in terms of frequency and pitch salience variations). Too short, lowsalience and harmonically related notes are then eliminated. Finally, the notes comprising the melody are extracted. This is the main topic of this paper. We select the melody notes by making use of note saliences and melodic smoothness. First, we select the notes with highest pitch salience at each moment. Then, by the melodic smoothness principle, we exploit the fact that tonal melodies are usually smooth. Thus, long music intervals indicate the presence of possibly erroneous notes, which are substituted by notes that smooth out the melodic contour. Finally, false positives in the extracted melody should be eliminated. To this end, we remove spurious notes that correspond to abrupt drops in note saliences or durations. Additionally, note clustering is conducted to further discriminate between true melody notes and false positives.